Identifying hot-spot pathways in fishery science and technology innovation through temporal heterogeneous graph neural networks Authors: Fengwei Zhang, Xiaolin Wang, Jing Hu Published: 2026 ABSTRACT AbstractAccurately identifying “hot-spot pathways” in fishery science and technology (S&T) innovation is critical for food security, economic development, and ecological sustainability. Traditional technology foresight methods struggle to capture complex, dynamic evolutionary patterns in S&T innovation networks. Drawing on Dosi’s conceptual framework of technological trajectories as domain-inspired design heuristics-whereby Dosi’s qualitative concepts provide structural guidance for model design rather than formal axioms that exhaustively capture the theoretical framework-this study proposes DTH-GNN (Documents-based Temporal Heterogeneous Graph Neural Network), integrating Graph Neural Networks with dynamic evolutionary analysis to identify potential hot-spot pathways. We construct a dynamic heterogeneous knowledge graph from multi-source data (2010-2024) encompassing 32,847 publications, 8,956 patents, and 1,856 projects. DTH-GNN combines an R-GCN-based heterogeneous encoder with a GRU-based temporal evolution module, achieving AUC = 0.934 and AP = 0.928 (after rigorous leakage assessment), significantly outperforming GCN, R-GCN, and EvolveGCN baselines. Information-theoretic analysis indicates that temporal features account for a substantial share of mutual information in link prediction (31.8%, 95% CI: [28.4%, 35.1%]), comparable to structural features (29.1%) and higher than attribute features (19.7%). Three high-potential pathways are identified and validated through expert evaluation (Krippendorffś α = 0.804): Smart Aquaculture, Green Seed Industry, and Ecological Fisheries. These findings provide data-driven scientific support for S&T investment prioritization in the fishery sector.Keywords: Fishery science and technology, Hot-spot pathways, Temporal heterogeneous graph, Graph neural networks, Mutual information analysis, Link prediction, Technology foresightSubject terms: Computational biology and bioinformatics, Ecology, Ecology, Environmental sciences, Mathematics and computing Introduction Fisheries constitute a vital primary industry underpinning national food security, regional economic development, and marine ecological balance1. Entering the third decade of the twenty-first century, global fisheries face unprecedented compound challenges: traditional capture fisheries suffer continuous resource decline owing to overfishing and climate change; intensive aquaculture brings environmental pollution, recurrent disease, and germplasm degradation; and consumers increasingly demand greater product quality, safety, and nutritional value, imposing ever-higher requirements for industry-wide modernization2,3. In this context, S&T innovation is widely regarded as the fundamental driving force for breaking through fishery development bottlenecks and leading industrial transformation. Fishery S&T innovation ecosystems are, however, complex, dynamic, and highly uncertain processes. It spans basic research, technology development, technology transfer, and market application, and involves multiple innovation actors-universities, research institutes, enterprises, and government agencies-operating collaboratively4. Within this complex innovation ecosystem, S&T resources are finite. Scientifically anticipating technological trends and precisely identifying and positioning “hot-spot pathways” with disruptive potential capable of driving industrial transformation has therefore become a major challenge for research managers and policymakers5. Traditional qualitative analyses and simple quantometric methods exhibit manifest limitations in addressing this challenge6. First, expert experience cannot process massive multi-source heterogeneous data and is prone to cognitive bias. Second, static text-mining methods fail to capture the dynamic evolutionary patterns of technology networks. Third, existing network-analysis approaches typically assume networks are homogeneous and static, ignoring both the heterogeneity and the time-varying nature of fishery innovation networks. These considerations motivate the introduction of frontier artificial intelligence technologies. Specifically, Graph Neural Networks (GNN) have become ideal tools for processing graph-structured data and capturing complex inter-entity relationships7,8. The development of temporal GNNs further makes it possible to simultaneously model network heterogeneity and dynamic evolution9,10. The present study transcends traditional qualitative analyses and simple quantometric approaches by incorporating frontier AI techniques to construct an analytical framework capable of objectively and dynamically revealing the internal logic and future directions of fishery S&T innovation. This not only enhances the scientific precision of S&T management but also stimulates collaborative innovation across industry, academia, and research, ultimately accelerating China’s fisheries toward a knowledge-intensive, resource-conserving, and environmentally friendly modern industrial paradigm. Literature review Identifying S&T hot-spot pathways requires methods that can simultaneously handle three challenges: data heterogeneity, temporal dynamics, and domain specificity. We review existing approaches through the lens of which challenges they address and which they leave unresolved, revealing a convergent gap. Bibliometric and patent-analysis methods represent the dominant paradigm for technology trend identification. Keyword co-occurrence analysis, citation analysis, and patent mapping have been extensively applied to depict the evolutionary landscape of technology domains6,11,12. However, these approaches share a fundamental limitation: they treat technologies as isolated units connected by co-occurrence statistics, ignoring the social-technical networks-projects, institutions, researchers-that produce and mediate technological innovation4,13. This is not merely an omission of additional data sources; it is a theoretical blind spot. As innovation systems theory emphasizes, knowledge creation and diffusion are relational processes occurring through networks of collaboration and citation among heterogeneous actors4. Methods that cannot model these relational structures are, by construction, unable to capture the cross-boundary knowledge flows that drive technological convergence. AI and machine learning approaches in the fishery domain have primarily addressed point-level technical problems-fish species recognition14, yield prediction15, and remote sensing classification16-rather than network-level strategic analysis. This gap is not coincidental but reflects a deeper methodological constraint: most AI models (CNNs, RNNs, ARIMA) operate on Euclidean data structures (images, time series), whereas innovation networks are inherently non-Euclidean graph structures. The recent emergence of GNNs-specifically designed for graph-structured data-opens a methodological pathway that was previously inaccessible7,8. Graph neural networks and temporal network analysis have achieved success in recommender systems, knowledge graph reasoning, and bioinformatics8,17,18. However, existing applications share two limitations relevant to our study. First, most temporal GNN architectures (e.g., EvolveGCN9, TGN10,19) model homogeneous graphs, collapsing the heterogeneous actor types that are theoretically critical for understanding innovation dynamics. Second, existing GNN-based technology forecasting studies13,20,21 treat pathway identification as a byproduct of graph embedding rather than as a formally defined prediction task with explicit theoretical grounding-consequently, their identified “pathways” are descriptive patterns rather than predictive constructs. Convergence gap The three streams converge on a common unsolved problem: no existing method simultaneously (a) models the heterogeneous multi-actor structure of innovation networks, (b) captures their temporal evolutionary dynamics, and (c) formalizes pathway identification as a theoretically grounded, predictive task. Addressing this convergence gap motivates the present study. Research gaps Based on the convergent limitations identified above, we specify three precise research gaps that the present study addresses: The analogy gap. Despite Dosi’s22 theoretical framework of technological trajectories being widely cited in innovation studies, no existing work has drawn inspiration from Dosi’s qualitative framework of technological trajectory to design a computationally tractable pathway extraction procedure that can be validated against real-world technology evolution data. Existing quantitative approaches either extract trajectories post-hoc from bibliometric data (losing predictive capacity) or apply GNNs as black-box predictors (losing theoretical grounding). Following Hesse23, we characterize our approach as an analogical mapping: Dosi’s concepts-technological paradigm, technological trajectory, and selection environment-provide structural guidance for model design, but do not constitute formal axioms that exhaustively capture the theoretical framework. Our Definition 1 addresses this gap by approximating Dosi’s qualitative intuition through four operational conditions that define a domain-inspired pathway extraction pipeline (Algorithm 1). We emphasize that this is an approximation-an analogical bridge between a rich sociological theory and a computationally tractable procedure-rather than an exhaustive formalization: Dosi’s constructs (e.g., “technological paradigm”) are complex socio-technical concepts that cannot be fully captured by graph-theoretic definitions. The analogy gap. Despite Dosi’s22 theoretical framework of technological trajectories being widely cited in innovation studies, no existing work has drawn inspiration from Dosi’s qualitative framework of technological trajectory to design a computationally tractable pathway extraction procedure that can be validated against real-world technology evolution data. Existing quantitative approaches either extract trajectories post-hoc from bibliometric data (losing predictive capacity) or apply GNNs as black-box predictors (losing theoretical grounding). Following Hesse23, we characterize our approach as an analogical mapping: Dosi’s concepts-technological paradigm, technological trajectory, and selection environment-provide structural guidance for model design, but do not constitute formal axioms that exhaustively capture the theoretical framework. Our Definition 1 addresses this gap by approximating Dosi’s qualitative intuition through four operational conditions that define a domain-inspired pathway extraction pipeline (Algorithm 1). We emphasize that this is an approximation-an analogical bridge between a rich sociological theory and a computationally tractable procedure-rather than an exhaustive formalization: Dosi’s constructs (e.g., “technological paradigm”) are complex socio-technical concepts that cannot be fully captured by graph-theoretic definitions. The heterogeneity-temporality integration gap. While heterogeneous GNNs (e.g., R-GCN8, HAN24) and temporal GNNs (e.g., EvolveGCN9, TGN10,19) have been developed independently, no architecture simultaneously leverages both heterogeneity and temporality for S&T pathway prediction. Our systematic review of the GNN literature (Web of Science Core Collection, query: TS=(“graph neural network” OR “temporal graph”) AND TS=(“technology forecasting” OR “innovation pathway” OR “emerging technology”), 2018-2024, n = 47 papers) confirms that no existing model addresses this integration gap. DTH-GNN fills this gap by combining R-GCN-based heterogeneous encoding with GRU-based temporal evolution. The heterogeneity-temporality integration gap. While heterogeneous GNNs (e.g., R-GCN8, HAN24) and temporal GNNs (e.g., EvolveGCN9, TGN10,19) have been developed independently, no architecture simultaneously leverages both heterogeneity and temporality for S&T pathway prediction. Our systematic review of the GNN literature (Web of Science Core Collection, query: TS=(“graph neural network” OR “temporal graph”) AND TS=(“technology forecasting” OR “innovation pathway” OR “emerging technology”), 2018-2024, n = 47 papers) confirms that no existing model addresses this integration gap. DTH-GNN fills this gap by combining R-GCN-based heterogeneous encoding with GRU-based temporal evolution. Relation to technology foresight methods We note that the present study is positioned within the computational technology forecasting tradition rather than the classical technology foresight tradition. Established foresight methodologies-including Delphi surveys25, technology roadmapping26, and patent trend analysis11-rely primarily on expert judgment or descriptive statistical patterns. These methods face two fundamental limitations when applied to pathway identification: (a) they cannot process the scale and heterogeneity of multi-source innovation data (publications, patents, projects, institutions), and (b) they do not provide predictive pathway construction as a formally defined computational task. DTH-GNN addresses these limitations by operationalizing pathway identification as a link prediction problem on a heterogeneous temporal graph, while acknowledging that its “predictions” are correlational rather than causal (see Limitations section). Classical foresight methods remain indispensable for qualitative context-setting, policy feasibility assessment, and stakeholder engagement-functions that fall outside the scope of our computational framework. (3)The domain-specific evidence gap. No prior study provides large-scale longitudinal evidence on the structural properties (degree distribution, convergence patterns, temporal clustering) of fishery S&T innovation networks. While such evidence exists for general technology domains6,20, the domain-specific characteristics of fishery innovation-such as the regulatory-driven adoption cycles, the strong role of government-funded projects, and the cross-domain convergence of ICT/biotechnology/marine science-produce distinct network dynamics that cannot be assumed from general-domain studies. Our 15-year longitudinal dataset (2010-2024) provides the first such evidence for the fishery sector. The domain-specific evidence gap. No prior study provides large-scale longitudinal evidence on the structural properties (degree distribution, convergence patterns, temporal clustering) of fishery S&T innovation networks. While such evidence exists for general technology domains6,20, the domain-specific characteristics of fishery innovation-such as the regulatory-driven adoption cycles, the strong role of government-funded projects, and the cross-domain convergence of ICT/biotechnology/marine science-produce distinct network dynamics that cannot be assumed from general-domain studies. Our 15-year longitudinal dataset (2010-2024) provides the first such evidence for the fishery sector. Research objectives and outline This study investigates computational model methods that effectively integrate multi-source heterogeneous data and capture dynamic evolutionary characteristics, with the goal of precisely identifying and predicting key hot-spot pathways in the fishery S&T domain. The main contributions are: Theoretical contribution. We draw on Dosi’s qualitative framework of technological trajectories as a domain-inspired design guide for constructing temporal heterogeneous graph representations, yielding two empirically testable hypotheses and one longitudinal empirical characterization: Increment 1 (theoretical-computational analogy). Definition 1approximates Dosi’s qualitative notion of technological trajectory through four operational conditions that define a domain-inspired pathway extraction procedure. Following Hesse23, we characterize this as an analogical mapping that provides structural guidance for model design rather than a formal specification that exhaustively captures Dosi’s theoretical framework. We stress that Dosi’s “technological paradigm” and “technological trajectory” are complex socio-technical constructs shaped by institutional incentives, cognitive frames, and economic selection pressures-none of which are directly encoded in our graph representation. Our operational conditions selectively capture the directionality and sequential dependency aspects of Dosi’s theory, while leaving other dimensions (e.g., paradigmatic lock-in mechanisms, cognitive frames of engineers) unmodeled. Unlike prior work that treats trajectory identification as post-hoc pattern extraction, our formulation is predictive by construction. Increment 2 (hypothesis derivation). We derive two empirically testable hypotheses (H1, H2) grounded in Dosi’s intuition that technological evolution is path-dependent, generating quantitative predictions testable against ablation experiments and data-splitting comparisons, closing the theory-experiment loop. This represents an advance over existing GNN-for-technology studies that employ GNNs as black-box predictors without explicit domain-theoretic motivation. Theoretical contribution. We draw on Dosi’s qualitative framework of technological trajectories as a domain-inspired design guide for constructing temporal heterogeneous graph representations, yielding two empirically testable hypotheses and one longitudinal empirical characterization: Increment 1 (theoretical-computational analogy). Definition 1approximates Dosi’s qualitative notion of technological trajectory through four operational conditions that define a domain-inspired pathway extraction procedure. Following Hesse23, we characterize this as an analogical mapping that provides structural guidance for model design rather than a formal specification that exhaustively captures Dosi’s theoretical framework. We stress that Dosi’s “technological paradigm” and “technological trajectory” are complex socio-technical constructs shaped by institutional incentives, cognitive frames, and economic selection pressures-none of which are directly encoded in our graph representation. Our operational conditions selectively capture the directionality and sequential dependency aspects of Dosi’s theory, while leaving other dimensions (e.g., paradigmatic lock-in mechanisms, cognitive frames of engineers) unmodeled. Unlike prior work that treats trajectory identification as post-hoc pattern extraction, our formulation is predictive by construction. Increment 2 (hypothesis derivation). We derive two empirically testable hypotheses (H1, H2) grounded in Dosi’s intuition that technological evolution is path-dependent, generating quantitative predictions testable against ablation experiments and data-splitting comparisons, closing the theory-experiment loop. This represents an advance over existing GNN-for-technology studies that employ GNNs as black-box predictors without explicit domain-theoretic motivation. Methodological contribution. The DTH-GNN model is developed as the first temporal heterogeneous GNN architecture specifically designed for S&T innovation network analysis, demonstrating that simultaneous modeling of network heterogeneity and temporal dynamics significantly improves pathway prediction accuracy (ΔAUC = +0.098 vs. R-GCN on the full graph under temporal split). Methodological contribution. The DTH-GNN model is developed as the first temporal heterogeneous GNN architecture specifically designed for S&T innovation network analysis, demonstrating that simultaneous modeling of network heterogeneity and temporal dynamics significantly improves pathway prediction accuracy (ΔAUC = +0.098 vs. R-GCN on the full graph under temporal split). Empirical contribution. The first large-scale longitudinal evidence (2010-2024) for power-law distribution and stage-wise convergence patterns in fishery S&T innovation networks is provided, extending innovation systems theory to domain-specific contexts. Empirical contribution. The first large-scale longitudinal evidence (2010-2024) for power-law distribution and stage-wise convergence patterns in fishery S&T innovation networks is provided, extending innovation systems theory to domain-specific contexts. Increment 3 (empirical discovery). Our analysis reveals that temporal features account for a substantial share of mutual information with link existence (31.8%, 95% CI: [28.4%, 35.1%]), comparable to structural features (29.1%) and higher than attribute features (19.7%). This provides descriptive quantitative evidence that temporal dynamics represent an important discriminative information source in fishery S&T pathway prediction, as further evaluated by the ablation experiment (Table 15). Table 15Validation results of different model variants in ablation experiments.Model configurationAUCAPRecallF1-scoreBaseline (GCN)0.7850.7600.7400.750+ TEM0.8450.8200.8100.815+ HGA0.8950.8750.8680.871+ DGU0.9100.8900.8820.886DTH-GNN (full)0.9340.9280.9050.911Each row incrementally adds one module to the configuration of the previous row: Baseline = GCN; +TEM = GCN + temporal encoding (GRU); +HGA = GCN + TEM + HGA (R-GCN), differs from standalone R-GCN; +DGU = GCN + TEM + HGA + dynamic graph update. DTH-GNN (Full) is trained end-to-end with all modules jointly optimized.Significant values are in bold. Validation results of different model variants in ablation experiments. Each row incrementally adds one module to the configuration of the previous row: Baseline = GCN; +TEM = GCN + temporal encoding (GRU); +HGA = GCN + TEM + HGA (R-GCN), differs from standalone R-GCN; +DGU = GCN + TEM + HGA + dynamic graph update. DTH-GNN (Full) is trained end-to-end with all modules jointly optimized. Significant values are in bold. Figure 1 illustrates the research technical roadmap. The study proceeds as follows: Construction of a dynamic heterogeneous knowledge graph of fishery S&T innovation. Multi-source data from 2010-2024, including S&T publications, patents, and national research projects, are integrated to extract core entities (technology keywords, projects, institutions, and researchers) and their interrelations, constructing a temporally evolving dynamic heterogeneous network. Construction of a dynamic heterogeneous knowledge graph of fishery S&T innovation. Multi-source data from 2010-2024, including S&T publications, patents, and national research projects, are integrated to extract core entities (technology keywords, projects, institutions, and researchers) and their interrelations, constructing a temporally evolving dynamic heterogeneous network. Development of the DTH-GNN model. A graph neural network model capable of processing heterogeneous nodes and edges while capturing their temporal dynamics is constructed to learn effective embedding representations of each entity at different time points. Development of the DTH-GNN model. A graph neural network model capable of processing heterogeneous nodes and edges while capturing their temporal dynamics is constructed to learn effective embedding representations of each entity at different time points. Implementation of hot-spot pathway identification and prediction. The hot-spot pathway identification problem is reformulated as a link prediction task on the graph, predicting the probability of strong connections forming between technology nodes in the future. Implementation of hot-spot pathway identification and prediction. The hot-spot pathway identification problem is reformulated as a link prediction task on the graph, predicting the probability of strong connections forming between technology nodes in the future. Empirical analysis and case validation. Using China’s fishery S&T development as the empirical object, the model is executed and its outputs analyzed in depth, with “smart fisheries” selected as a case study to validate effectiveness and provide policy recommendations. Empirical analysis and case validation. Using China’s fishery S&T development as the empirical object, the model is executed and its outputs analyzed in depth, with “smart fisheries” selected as a case study to validate effectiveness and provide policy recommendations. Research technical roadmap of the proposed framework. Related work Theory of hot-spot pathways Theoretical foundations: from technological trajectories to network-based pathway identification Understanding the dynamics of technological evolution has been a central concern in innovation studies. Dosi formalized the concept of technological paradigms and technological trajectories within a paradigm as the “paths of ‘normal’ problem solving activity” guided by the paradigm’s “implicit or explicit ‘model’ and ‘exemplar’ ”22. Nelson and Winter further developed an evolutionary theory of economic change, emphasizing the routines, search behaviors, and selection environment that shape technological development over time27. Building upon these foundational works, contemporary innovation systems theory4 extends the analytical lens from individual firms to the broader ecosystem of actors, institutions, and interactions that co-produce technological change. A critical insight from innovation systems theory is that knowledge creation and diffusion are fundamentally relational processes-they occur through networks of collaboration, citation, and co-invention among heterogeneous actors. This relational nature suggests that technological evolution can be inspired through network structures. Specifically, we propose that: The nodes of an innovation network represent technological concepts, actors, and artifacts, whose properties (centrality, connectivity patterns) encode their role in the technological ecosystem. The nodes of an innovation network represent technological concepts, actors, and artifacts, whose properties (centrality, connectivity patterns) encode their role in the technological ecosystem. The edges of an innovation network represent relationships of various types (collaboration, citation, co-occurrence), whose weights and temporal dynamics capture the strength and evolution of technological connections. The edges of an innovation network represent relationships of various types (collaboration, citation, co-occurrence), whose weights and temporal dynamics capture the strength and evolution of technological connections. Pathways-ordered sequences of technologically and economically related nodes-represent our domain-inspired approximation of Dosi’s technological trajectories within a modern, data-rich environment where network representations are computationally tractable. We use the term “domain-inspired approximation” to signal that this operational definition captures the sequential and directional aspects of Dosi’s trajectories while abstracting from the broader socio-institutional context (e.g., firm-level capabilities, regulatory environments) that Dosi’s original framework encompasses. Pathways-ordered sequences of technologically and economically related nodes-represent our domain-inspired approximation of Dosi’s technological trajectories within a modern, data-rich environment where network representations are computationally tractable. We use the term “domain-inspired approximation” to signal that this operational definition captures the sequential and directional aspects of Dosi’s trajectories while abstracting from the broader socio-institutional context (e.g., firm-level capabilities, regulatory environments) that Dosi’s original framework encompasses. Definition of hot-spot pathways in fishery S&T innovation Building upon the theoretical foundations established above, we now provide a formal definition of hot-spot pathways tailored to the fishery science and technology (S&T) innovation context. Definition 1 (Hot-Spot Pathway). A hot-spot pathway in fishery S&T innovation is defined as an ordered sequence of technology nodes where (k = 9 corresponds to L = 8 edges in Algorithm 1), constructed by Algorithm 1, satisfying the following operational conditions that are jointly necessary and sufficient: (i)Link Predictability (→ Algorithm 1, Step 6). For any consecutive pair , the DTH-GNN model assigns a predicted link probability for some observed graph , where τ = 0.50 is the calibrated confidence threshold (selected by F1-maximization on the 2023 validation set; see Algorithm 2), and is computed using the node embeddings learned from via the bilinear decoder (Eq. 13). Link Predictability (→ Algorithm 1, Step 6). For any consecutive pair , the DTH-GNN model assigns a predicted link probability for some observed graph , where τ = 0.50 is the calibrated confidence threshold (selected by F1-maximization on the 2023 validation set; see Algorithm 2), and is computed using the node embeddings learned from via the bilinear decoder (Eq. 13). The pathway construction in Algorithm 1 operates on the final observed graph snapshot , treating the prediction target as the next time step t+1 relative to this observation point. The node embeddings at each time step t encode historical evolution patterns through the GRU module, enabling the model to capture temporal dynamics while operating on a fixed observed graph for computational tractability. Condition (i) requires that every edge in a pathway exhibits above-random predictability. Because the positive-to-negative edge ratio in our dataset is 1:1, random-chance link probability is exactly 0.50. Condition (i) therefore implements the requirement that every included edge must be distinguishable from random chance - a stricter requirement than merely being non-zero. The specific value τ = 0.50 and its selection rationale are detailed in condition (iv) and Algorithm 2. (ii)Greedy sequential extension (→ Algorithm 1, Steps 6-7). Each node is selected as the highest-probability neighbor of among all candidates not already in P, i.e., 1 Greedy sequential extension (→ Algorithm 1, Steps 6-7). Each node is selected as the highest-probability neighbor of among all candidates not already in P, i.e., 1 (iii)Pathway strength (→ Algorithm 1, Step 14). The pathway score is the geometric mean of constituent edge probabilities: 2 providing a continuous ranking metric for comparing candidate pathways of different lengths. Pathway strength (→ Algorithm 1, Step 14). The pathway score is the geometric mean of constituent edge probabilities: 2 providing a continuous ranking metric for comparing candidate pathways of different lengths. (iv)Confidence termination (→ Algorithm 1, Steps 8-10). Pathway extension terminates when the maximum candidate probability falls below a calibrated threshold τ (τ = 0.50, selected by F1-maximization on the validation set; see Algorithm 2). This condition ensures that only edges with sufficient predictive confidence are included, distinguishing “hot-spot” pathways from routine evolutionary sequences. Conditions (i)-(iv) are not independent design choices but are jointly entailed by the computational procedure of Algorithm 1. Condition (i) establishes the prerequisite that all edges be predictable; condition (ii) specifies the selection mechanism; condition (iii) defines the aggregation rule for pathway-level scoring; condition (iv) provides the termination criterion. Together, they draw on Dosi’s intuition about directionality in technological problem-solving to motivate a computationally tractable procedure for pathway extraction. However, we emphasize a critical scope limitation: these conditions operationalize only the structural and temporal dimensions of Dosi’s framework-the “what connects to what” and “when” questions-while leaving unaddressed the causal mechanisms (“why do trajectories form?”) and the socio-institutional dynamics (“who drives trajectory selection?”) that constitute central concerns in Dosi’s original theory22. The correlational nature of our framework is further discussed in the Limitations section and the epistemological clarification below. Confidence termination (→ Algorithm 1, Steps 8-10). Pathway extension terminates when the maximum candidate probability falls below a calibrated threshold τ (τ = 0.50, selected by F1-maximization on the validation set; see Algorithm 2). This condition ensures that only edges with sufficient predictive confidence are included, distinguishing “hot-spot” pathways from routine evolutionary sequences. Conditions (i)-(iv) are not independent design choices but are jointly entailed by the computational procedure of Algorithm 1. Condition (i) establishes the prerequisite that all edges be predictable; condition (ii) specifies the selection mechanism; condition (iii) defines the aggregation rule for pathway-level scoring; condition (iv) provides the termination criterion. Together, they draw on Dosi’s intuition about directionality in technological problem-solving to motivate a computationally tractable procedure for pathway extraction. However, we emphasize a critical scope limitation: these conditions operationalize only the structural and temporal dimensions of Dosi’s framework-the “what connects to what” and “when” questions-while leaving unaddressed the causal mechanisms (“why do trajectories form?”) and the socio-institutional dynamics (“who drives trajectory selection?”) that constitute central concerns in Dosi’s original theory22. The correlational nature of our framework is further discussed in the Limitations section and the epistemological clarification below. To clarify the epistemological status of this mapping, we distinguish two levels of scientific reasoning23,28: Theoretical inspiration (analogical reasoning): Dosi’s concepts serve as design heuristics, not formal axioms. Specifically:Dosi’s “technological paradigm” → our heterogeneous node taxonomy (T/P/INS/PER/LIT);Dosi’s “technological trajectory” → our pathway extraction procedure (Eqs. 1–2);Dosi’s “selection environment” → our ENR metric (Table 7). This analogical mapping provides structural guidance for model design, following established practices in computational social science29. Theoretical inspiration (analogical reasoning): Dosi’s concepts serve as design heuristics, not formal axioms. Specifically:Dosi’s “technological paradigm” → our heterogeneous node taxonomy (T/P/INS/PER/LIT);Dosi’s “technological trajectory” → our pathway extraction procedure (Eqs. 1–2);Dosi’s “selection environment” → our ENR metric (Table 7). This analogical mapping provides structural guidance for model design, following established practices in computational social science29. Dosi’s “technological paradigm” → our heterogeneous node taxonomy (T/P/INS/PER/LIT); Dosi’s “technological paradigm” → our heterogeneous node taxonomy (T/P/INS/PER/LIT); Dosi’s “technological trajectory” → our pathway extraction procedure (Eqs. 1–2); Dosi’s “technological trajectory” → our pathway extraction procedure (Eqs. 1–2); Dosi’s “selection environment” → our ENR metric (Table 7). Dosi’s “selection environment” → our ENR metric (Table 7). Empirical validation (correlational inference): hypotheses H1 and H2 are validated through three empirical mapping steps:Step 1 (pathway → edge): pathway strength (Eq. 2) decomposes into edge probabilities ;Step 2 (predictability → prediction): condition (i) maps to link prediction: given , predict edge existence at ;Step 3 (termination → threshold): condition (iv) corresponds to confidence threshold τ, calibrated by F1-maximization on the 2023 validation set. Reporting AUC/F1 establishes that DTH-GNN accurately traces observed technology co-occurrence patterns (“does the model work?”), but does not establish causal mechanisms (“why do trajectories form?”). Empirical validation (correlational inference): hypotheses H1 and H2 are validated through three empirical mapping steps:Step 1 (pathway → edge): pathway strength (Eq. 2) decomposes into edge probabilities ;Step 2 (predictability → prediction): condition (i) maps to link prediction: given , predict edge existence at ;Step 3 (termination → threshold): condition (iv) corresponds to confidence threshold τ, calibrated by F1-maximization on the 2023 validation set. Reporting AUC/F1 establishes that DTH-GNN accurately traces observed technology co-occurrence patterns (“does the model work?”), but does not establish causal mechanisms (“why do trajectories form?”). Step 1 (pathway → edge): pathway strength (Eq. 2) decomposes into edge probabilities ; Step 1 (pathway → edge): pathway strength (Eq. 2) decomposes into edge probabilities ; Step 2 (predictability → prediction): condition (i) maps to link prediction: given , predict edge existence at ; Step 2 (predictability → prediction): condition (i) maps to link prediction: given , predict edge existence at ; Step 3 (termination → threshold): condition (iv) corresponds to confidence threshold τ, calibrated by F1-maximization on the 2023 validation set. Step 3 (termination → threshold): condition (iv) corresponds to confidence threshold τ, calibrated by F1-maximization on the 2023 validation set. Two additional theoretical perspectives strengthen the link between the computational model and innovation theory: Knowledge recombination theory30: Fleming’s model explains why co-occurrence patterns carry predictive information: new technologies emerge from novel combinations of existing knowledge elements. Our heterogeneous graph captures these recombination dynamics through multi-relational edges encoding collaborative, institutional, and citation-mediated connections. Knowledge recombination theory30: Fleming’s model explains why co-occurrence patterns carry predictive information: new technologies emerge from novel combinations of existing knowledge elements. Our heterogeneous graph captures these recombination dynamics through multi-relational edges encoding collaborative, institutional, and citation-mediated connections. Technological complementarity31: Complementarity explains why certain co-occurrence patterns are more persistent and predictable. DTH-GNN’s bilinear decoder captures asymmetric complementarity effects through the learnable relation matrix . Technological complementarity31: Complementarity explains why certain co-occurrence patterns are more persistent and predictable. DTH-GNN’s bilinear decoder captures asymmetric complementarity effects through the learnable relation matrix . Table 7Statistical characteristics of the fishery technology collaboration network (2010–2024).YearTech. nodesProj. nodesInst. nodesRes. nodesLit. nodesTotal nodesTotal edgesENR1,245894232,1562,1076,02045,8727.6220111,6411235192,7552,8327,87061,0727.7620121,8701566123,3013,5579,49677,2028.1320132,1981896953,8344,28311,19994,2968.4220142,4342187834,2825,00812,725112,3208.832,8472548674,8925,73314,593128,9348.8420164,6263561,0076,0057,27619,270182,4769.4620175,7743881,1166,7778,81922,874240,17710.5020186,3724211,2047,65410,36126,012305,38111.7420197,6085101,3598,40211,90429,783377,53512.678,9235681,4569,23413,44733,628456,12213.56202114,2808051,65411,00118,55146,291624,50613.49202218,8299831,78512,29723,65457,548793,01113.78202324,9931,2362,01114,29328,75871,291962,42913.5029,8761,4312,15615,67833,86283,0031,132,54713.64ENR (edge-to-node ratio) = total edges / total nodes. Statistical characteristics of the fishery technology collaboration network (2010–2024). ENR (edge-to-node ratio) = total edges / total nodes. Limitations This correlational framework has three methodological limitations: (1) Confounding: two technologies may co-occur frequently without direct causal influence; (2) Selection bias: our dataset prioritizes publicly funded research (CNKI, CNPD, NSFC), potentially underrepresenting private-sector innovations; (3) Causal identification: establishing causal effects would require exogenous shocks (e.g., policy mandates as instrumental variables). These constitute important directions for future work. Empirically motivated hypotheses Based on the domain-inspired design framework, we derive two empirically motivated hypotheses that guide our experimental design. We deliberately use the term “empirically motivated” rather than “theoretical” to signal that these hypotheses are engineering design choices informed by domain knowledge rather than logical deductions from Dosi’s formal theory. Dosi’s framework motivates why one might expect heterogeneity and temporality to matter for pathway prediction, but it does not logically entail that R-GCN + GRU is the optimal architecture or that AUC improvements necessarily confirm the theory. These hypotheses generate performance predictions testable against ablation experiments and data-splitting comparisons: H1 (heterogeneity value-added hypothesis)[empirically motivated]. Predictions derived from heterogeneous graphs that explicitly distinguish actor types (technology, project, institution, researcher, literature) and relation types (collaboration, citation, funding, co-occurrence, containment, affiliation, undertaking) achieve significantly higher link prediction accuracy than predictions from homogeneous graphs that collapse these distinctions, with the improvement being primarily attributable to the preservation of cross-type information pathways (e.g., literature→technology, institution→project→technology) that mediate indirect technology-technology associations. This hypothesis generates the following testable prediction: removing heterogeneous relation types from DTH-GNN (i.e., reducing to homogeneous GCN) will cause a disproportionate accuracy decline for technology pairs that are connected primarily through multi-hop heterogeneous paths rather than direct co-occurrence edges. We operationalize this prediction in the ablation experiment (Table 15): the +HGA module should contribute more to AUC for edge types with high heterogeneous-path dependency than for edges directly connected by co-occurrence. H2 (temporal dynamics primacy hypothesis)[empirically motivated]. The temporal ordering of network evolution events provides predictive information that cannot be recovered from time-aggregated static features, because the directionality and recency of technology co-occurrence patterns encode path-dependent and regime-specific dynamics that static aggregation irreversibly destroys. This hypothesis generates the following testable prediction: replacing the temporal split with a random split will inflate static model performance more than temporal model performance, because static models exploit training-set diversity in ways that do not reflect genuine temporal generalization capacity to compensate for their inability to model temporal dynamics. Equivalently, the performance gap between DTH-GNN and static baselines should be significantly larger under temporal split than under random split. We test this prediction in the data splitting strategy comparison (Table 14). Table 14Performance comparison under temporal versus random splitting strategies.ModelSplit strategyAUCAPAUC inflation (random − temporal)DTH-GNN (proposed)Temporal0.9340.928–DTH-GNN (proposed)Random0.9410.935+0.007GCNTemporal0.7850.760–GCNRandom0.8920.875+0.107R-GCNTemporal0.8360.815–R-GCNRandom0.9010.882+0.065EvolveGCNTemporal0.8910.880–EvolveGCNRandom0.9050.895+0.014 “AUC Inflation (Random − Temporal)” measures the performance increase when temporal structure is removed. Smaller values indicate greater robustness to temporal shift and less reliance on temporal leakage. DTH-GNN’s smallest gain (+0.7%) is consistent with its most comprehensive temporal modeling design. The temporal-split AUC (0.934) is the reported main result, as it reflects realistic forecasting conditions where future data is unavailable. Performance comparison under temporal versus random splitting strategies. “AUC Inflation (Random − Temporal)” measures the performance increase when temporal structure is removed. Smaller values indicate greater robustness to temporal shift and less reliance on temporal leakage. DTH-GNN’s smallest gain (+0.7%) is consistent with its most comprehensive temporal modeling design. The temporal-split AUC (0.934) is the reported main result, as it reflects realistic forecasting conditions where future data is unavailable. Scope note: H1 and H2 are labeled “empirically motivated” to distinguish them from formal theoretical deductions. While Dosi’s framework motivates why heterogeneity and temporality might matter for pathway prediction, it does not logically entail that R-GCN+GRU is the optimal architecture. H1/H2 therefore represent engineering design hypotheses informed by Dosi’s intuition about path-dependence and technology selection environments, rather than theorems derived from Dosi’s axioms. Graph neural networks (GNN) A Graph Neural Network is a deep learning model specifically designed for processing and learning from graph-structured data. A graph G = (V, E) consists of a set of vertices V and a set of edges E. The core idea of GNN is message passing32: each node updates its representation (embedding) by aggregating information from its neighbors. This process typically iterates over multiple rounds (layers), enabling each node to capture multi-hop neighborhood structural information within the graph. A typical GNN layer can be expressed as Eq. (3): where is the feature representation of node v at layer l, is the set of neighbor nodes of v, the AGGREGATE function pools neighbor information, and UPDATE combines aggregated information with the node’s own information to generate the new representation. Popular models such as Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) differ primarily in their specific implementations of AGGREGATE and UPDATE. The power of GNN lies in its end-to-end learning of node feature representations that encode rich topological structure information, making it well suited for downstream tasks such as node classification and link prediction33. Temporal graph neural networks Many real-world networks-social networks, knowledge graphs, and others-evolve dynamically over time. Fishery S&T innovation networks are no exception: new projects are continuously established, new publications and patents emerge, and research collaborations change constantly. To capture this dynamism, it is necessary to introduce the concept of dynamic or temporal graphs. Temporal GNNs (TGNs) extend GNNs to dynamic graphs, with the core objective of learning an evolution function f(v, t) whose output captures the state of node v at time t. Depending on the treatment of time, TGNs fall into two broad categories: (i) discrete-time models, which partition time into a series of snapshots and apply GNN to each snapshot while using RNN, LSTM, or Transformer-type sequence models to capture temporal dependencies across snapshots9,34; and (ii) continuous-time models, which treat graph evolution as an event stream and directly model timestamped events using temporal point processes or temporal encoding functions10,19. This study adopts the discrete-time model, as it aligns naturally with the annual-cycle release pattern of S&T outputs (publications, patents, and projects) and facilitates data processing and model construction. Recent advances in temporal graph learning have introduced transformer-based architectures for dynamic graph representation. Among these, DyFormer35 represents the state-of-the-art temporal graph Transformer for link prediction, which we evaluate on a restricted subgraph due to its attention complexity. Notably, GraphGPS36 is designed for static graph representation learning and does not natively support temporal dynamics, making it fundamentally inapplicable to temporal link prediction tasks. The R-GCN + GRU architecture in DTH-GNN represents a favorable trade-off between heterogeneous temporal modeling and computational efficiency for S&T innovation network analysis. Heterogeneous graph neural networks (HGN) Traditional GNNs assume homogeneous graphs in which all nodes and edges share a single type. However, fishery S&T innovation networks are typical heterogeneous networks containing multiple node types-technology, project, institution, researcher, and literature-as well as multiple edge types, including collaboration, citation, and funding. Applying a homogeneous GNN to such a network ignores node and edge type information, leading to information loss. Heterogeneous GNNs (HGNs) are designed precisely to address this issue. Their core idea is to aggregate information based on meta-paths or relation types: for an “institution” node, information from “collaborating” “institution” neighbors and from “funded” “project” neighbors is aggregated separately. Different types of neighbor information are processed through distinct transformation matrices, thereby preserving the heterogeneous semantic information of the network17. This study adopts the R-GCN architecture8 for handling heterogeneous relations. R-GCN is a classical model for heterogeneous graphs that learns an independent weight matrix for each relation type and selects the corresponding matrix during message passing. R-GCN directly parameterizes relation types, which aligns with the seven explicitly defined edge relations in our heterogeneous graph (collaboration, affiliation, undertaking, funding, citation, co-occurrence, and containment), while maintaining a controllable parameter count suitable for large-scale training on data with millions of edges. Alternative heterogeneous GNN architectures include HAN24 and HetGNN37, which employ attention mechanisms and random walk-based sampling, respectively. While these models achieve competitive performance on benchmark datasets, R-GCN’s explicit relation parameterization aligns more closely with the predefined edge typology (seven relation types in Table 6) and offers more interpretable attention weights for domain analysis. Table 6Definition and weighting of edges in the dynamic heterogeneous graph.RelationEdge structureEdge definitionWeightCollaboration(R)-[collaborates with]-(R)Two researchers co-author a publication, co-file a patent, or jointly participate in a projectAnnual collaboration countAffiliation(R)-[affiliated with]-(I)Relation between researcher and their institutionFixed at 1Undertaking(P)-[undertaken by]-(I)Institution undertakes a research projectFixed at 1Funding(P)-[funds]-(R)Researcher serves as principal investigator or key memberFixed at 1Citation(A)-[cites]-(A)Citation relation between S&T documents; key to constructing technology inheritance and evolution pathsFixed at 1Co-occurrence(T)-[co-occurs with]-(T)Two technology keywords appear together in the same publication, patent, or projectAnnual co-occurrence frequency, weighted by PMIContainment / projected association(A/P)-[contains]-(T); (I/R)-[associated with]-(T)A/P-T edges are direct text-derived keyword-containment edges from titles, abstracts, and keyword fields. I/R-T edges are projected associations derived from institution or researcher links to A/P records that contain the technology keyword.Direct A/P-T: TF-IDF; projected I/R-T: aggregated TF-IDF over linked records Definition and weighting of edges in the dynamic heterogeneous graph. Materials and methods Overall framework To achieve systematic identification of hot-spot pathways in fishery S&T projects, this study designs a four-stage analytical framework (Fig. 2). Data layer: multi-source data-publications, patents, and projects from 2010 to 2024-are collected from CNKI, Wanfang Data, the National Natural Science Foundation of China (NSFC) project retrieval system, Derwent Innovation Index, and CNIPA patent databases. Data layer: multi-source data-publications, patents, and projects from 2010 to 2024-are collected from CNKI, Wanfang Data, the National Natural Science Foundation of China (NSFC) project retrieval system, Derwent Innovation Index, and CNIPA patent databases. Graph construction layer: raw data undergo cleaning, entity extraction (e.g., extracting authors, institutions, and keywords from publications), relation construction (e.g., co-authorship, citation, funding), and temporal slicing, yielding a series of annual dynamic heterogeneous graph snapshots . Graph construction layer: raw data undergo cleaning, entity extraction (e.g., extracting authors, institutions, and keywords from publications), relation construction (e.g., co-authorship, citation, funding), and temporal slicing, yielding a series of annual dynamic heterogeneous graph snapshots . Model learning layer: the dynamic heterogeneous graph sequence is fed into the DTH-GNN model, which learns end-to-end to generate a low-dimensional dense embedding vector for each node at each time step, encoding the node’s attributes and structural/semantic information within the dynamic network. Model learning layer: the dynamic heterogeneous graph sequence is fed into the DTH-GNN model, which learns end-to-end to generate a low-dimensional dense embedding vector for each node at each time step, encoding the node’s attributes and structural/semantic information within the dynamic network. Pathway identification and application layer: hot-spot pathway identification is modeled as a link prediction task between technology nodes. Learned node embeddings are used to compute the probability of a connection forming between any two technology nodes at future time step t+1. High-probability connections are treated as potential innovation opportunity points; connecting these points in temporal order yields candidate hot-spot pathways, which are subsequently interpreted, validated, and visualized with domain expert knowledge. Pathway identification and application layer: hot-spot pathway identification is modeled as a link prediction task between technology nodes. Learned node embeddings are used to compute the probability of a connection forming between any two technology nodes at future time step t+1. High-probability connections are treated as potential innovation opportunity points; connecting these points in temporal order yields candidate hot-spot pathways, which are subsequently interpreted, validated, and visualized with domain expert knowledge. The proposed four-stage framework for hot-spot pathway identification. Data preparation and preprocessing Data sources The multi-source heterogeneous data used in this study comprise three categories: (1) Academic publications: journal articles and graduate dissertations published between January 2010 and December 2024, retrieved from CNKI and Wanfang Database using the topic terms “fishery”, “aquatic”, “fishing”, and “aquaculture”, along with titles, abstracts, keywords, authors, institutions, publication year, and references. (2) Invention patents: fishery-related invention patents (IPC code A01K) retrieved from the Derwent Innovation Index, with patent title, abstract, inventors, applicant (institution), and application year. (3) Research projects: national fishery-related research project data obtained from the NSFC project retrieval system and the National S&T Management Information System Public Service Platform, including project title, principal investigator, host institution, abstract, keywords, start/end year, and funding amount. Data collection Data collection followed the PRISMA-Emerging38 protocol, adapted for observational database studies. Adaptation rationale: while PRISMA was originally developed for systematic reviews in health sciences, its core principles-transparent reporting, explicit eligibility criteria, and reproducible selection procedures-are domain-agnostic and have been increasingly adopted in interdisciplinary research. The PRISMA extension for scoping reviews (PRISMA-ScR)39 demonstrates that the PRISMA framework can be validly extended beyond clinical medicine to broader evidence synthesis contexts. For our fishery S&T domain, we retain PRISMA-Emerging items applicable to multi-source data compilation (search strategy, deduplication protocol, inclusion/exclusion criteria) and omit health-specific items (e.g., clinical outcomes, risk of bias). Figure 3 illustrates our adapted flow diagram; detailed search strings and screening protocols are described in “Data collection protocol checklist”. PRISMA 2020 flow diagram for data collection. The screening proceeded as follows: For academic publications, we applied the same search query TS=(fishery OR aquatic OR fishing OR aquaculture OR marine fishery) AND (innovation OR technology OR science) in CNKI and Wanfang, covering 1 January 2010 to 31 December 2024; CNKI returned 48,692 records and Wanfang returned 41,611 records. For patents, IPC code A01K combined with relevant keywords in the Derwent Innovation Index produced 15,847 records. For projects, NSFC and the National S&T Management Platform were queried using “fishery” and “aquatic” keywords, yielding 3,256 national research projects. The total raw retrieval across all four sources was 109,406 records. Records were deduplicated in three steps. Step 0 identified 30,265 CNKI-Wanfang overlapping records by DOI match and removed them as a batch. Step 1 matched remaining records by unique identifier (DOI for publications, patent number for patents, project approval number for NSFC projects), removing 12,452 duplicates. Step 2 applied normalized Levenshtein string similarity (threshold ≥ 0.90) on title fields to catch cross-database duplicates without standard identifiers, removing an additional 1,244 records. After all three deduplication steps (total removed: 43,961), 65,445 unique records remained for screening (Table 1). Table 1Detailed breakdown of exclusion decisions at each screening stage.Screening phaseEligibility phase Exclusion criterionRecordsWithin-phase shareRecordsWithin-phase shareE1 (off-topic)465740.8%357734.5%E2 (incomplete metadata)676659.2%311030.0%E3 (non-invention patent)––367635.5%E4 (out-of-time)00%00%Total excluded11,423100%10,363100% Detailed breakdown of exclusion decisions at each screening stage. Screening phase (title and abstract): the 65,445 deduplicated records were first screened at title and abstract level. Records were excluded if they were clearly off-topic (E1) based on title and abstract, or if metadata (title, abstract, keywords) were incomplete (E2). A total of 11,423 records were excluded at the screening phase (E1: 4,657; E2: 6,766), leaving 54,022 records to proceed to full-text assessment. Eligibility phase (full-text assessment): the 54,022 records were assessed for eligibility based on the full record content. Records were excluded if they remained off-topic after full review (E1: 3,577), if metadata were still incomplete upon full inspection (E2: 3,110), or if they were non-invention patents (E3: 3,676). No records were excluded for being out-of-time (E4: 0), as the publication/application year was already constrained during the search. A total of 10,363 records were excluded at the eligibility phase. The four exclusion criteria were: E1—off-topic (not directly related to fishery S&T innovation, e.g., fishery economics or pure biology without technical content); E2—incomplete metadata (missing title, abstract, or keywords); E3—non-invention patent (utility models and design patents excluded); E4—out-of-time (publication or project year outside 2010–2024). The final dataset comprises 32,847 publications, 8956 invention patents, and 1856 research projects (total 43,659 records). Records were retained if they met all of the following conditions: (a) direct relevance to fishery S&T innovation as determined by two independent annotators (Cohen’s κ = 0.87); (b) publication/grant date between 2010 and 2024; and (c) complete title, abstract, and keyword information. Title and abstract screening were conducted independently by two trained research assistants with backgrounds in fishery science; discrepancies (11.4% of records, corresponding to approximately 7461 of the 65,445 post-deduplication records) were resolved through consultation with a senior domain expert. We assessed data saturation by randomly sampling 500 records from the excluded pool and re-evaluating them against inclusion criteria; the re-evaluation confirmed that excluded records lacked direct relevance to fishery S&T innovation, indicating that the screening process did not inadvertently exclude relevant records. Data collection protocol checklist: Search strings for each databaseCNKI (China National Knowledge Infrastructure). Advanced Search: TS = (“fishery” OR “aquatic product” OR “fishing” OR “aquaculture” OR “marine fishery” OR “fisheries science”) AND TS = (“innovation” OR “technology” OR “science & technology” OR “R&D” OR “research”) Filters: Publication date = 2010-01-01 to 2024-12-31; Document type = journal article, dissertation; Language = Chinese or English.Wanfang Database. Advanced Search (same Boolean logic as CNKI): (Title/Abstract/Keywords: fishery OR aquatic OR fishing OR aquaculture OR marine fishery) AND (Title/Abstract/Keywords: innovation OR technology OR “S&T” OR R&D); Year = 2010-2024; Media type = journal, dissertation.Derwent innovation index (clarivate). IPC = A01K* AND (TI = (fishery OR aquaculture OR “fish farming” OR mariculture)) AND (AB = (innovation OR technology OR technique OR method)); Patent publication date = 2010-2024; Patent type = invention.NSFC project retrieval system. Project name/abstract contains: “fishery” OR “aquatic” OR “marine fisheries” OR “aquaculture” OR “fish breeding”; Funding period overlapping 2010-2024; Project status = completed or ongoing.National S&T management information system public service platform. Project contains: “fishery” OR “aquatic products” OR “marine fishery” OR “aquaculture technology”; Year = 2010-2024; Project type = national key R&D, NSFC, “948” project, agricultural S&T key project. Search strings for each databaseCNKI (China National Knowledge Infrastructure). Advanced Search: TS = (“fishery” OR “aquatic product” OR “fishing” OR “aquaculture” OR “marine fishery” OR “fisheries science”) AND TS = (“innovation” OR “technology” OR “science & technology” OR “R&D” OR “research”) Filters: Publication date = 2010-01-01 to 2024-12-31; Document type = journal article, dissertation; Language = Chinese or English.Wanfang Database. Advanced Search (same Boolean logic as CNKI): (Title/Abstract/Keywords: fishery OR aquatic OR fishing OR aquaculture OR marine fishery) AND (Title/Abstract/Keywords: innovation OR technology OR “S&T” OR R&D); Year = 2010-2024; Media type = journal, dissertation.Derwent innovation index (clarivate). IPC = A01K* AND (TI = (fishery OR aquaculture OR “fish farming” OR mariculture)) AND (AB = (innovation OR technology OR technique OR method)); Patent publication date = 2010-2024; Patent type = invention.NSFC project retrieval system. Project name/abstract contains: “fishery” OR “aquatic” OR “marine fisheries” OR “aquaculture” OR “fish breeding”; Funding period overlapping 2010-2024; Project status = completed or ongoing.National S&T management information system public service platform. Project contains: “fishery” OR “aquatic products” OR “marine fishery” OR “aquaculture technology”; Year = 2010-2024; Project type = national key R&D, NSFC, “948” project, agricultural S&T key project. CNKI (China National Knowledge Infrastructure). Advanced Search: TS = (“fishery” OR “aquatic product” OR “fishing” OR “aquaculture” OR “marine fishery” OR “fisheries science”) AND TS = (“innovation” OR “technology” OR “science & technology” OR “R&D” OR “research”) Filters: Publication date = 2010-01-01 to 2024-12-31; Document type = journal article, dissertation; Language = Chinese or English. CNKI (China National Knowledge Infrastructure). Advanced Search: TS = (“fishery” OR “aquatic product” OR “fishing” OR “aquaculture” OR “marine fishery” OR “fisheries science”) AND TS = (“innovation” OR “technology” OR “science & technology” OR “R&D” OR “research”) Filters: Publication date = 2010-01-01 to 2024-12-31; Document type = journal article, dissertation; Language = Chinese or English. Wanfang Database. Advanced Search (same Boolean logic as CNKI): (Title/Abstract/Keywords: fishery OR aquatic OR fishing OR aquaculture OR marine fishery) AND (Title/Abstract/Keywords: innovation OR technology OR “S&T” OR R&D); Year = 2010-2024; Media type = journal, dissertation. Wanfang Database. Advanced Search (same Boolean logic as CNKI): (Title/Abstract/Keywords: fishery OR aquatic OR fishing OR aquaculture OR marine fishery) AND (Title/Abstract/Keywords: innovation OR technology OR “S&T” OR R&D); Year = 2010-2024; Media type = journal, dissertation. Derwent innovation index (clarivate). IPC = A01K* AND (TI = (fishery OR aquaculture OR “fish farming” OR mariculture)) AND (AB = (innovation OR technology OR technique OR method)); Patent publication date = 2010-2024; Patent type = invention. Derwent innovation index (clarivate). IPC = A01K* AND (TI = (fishery OR aquaculture OR “fish farming” OR mariculture)) AND (AB = (innovation OR technology OR technique OR method)); Patent publication date = 2010-2024; Patent type = invention. NSFC project retrieval system. Project name/abstract contains: “fishery” OR “aquatic” OR “marine fisheries” OR “aquaculture” OR “fish breeding”; Funding period overlapping 2010-2024; Project status = completed or ongoing. NSFC project retrieval system. Project name/abstract contains: “fishery” OR “aquatic” OR “marine fisheries” OR “aquaculture” OR “fish breeding”; Funding period overlapping 2010-2024; Project status = completed or ongoing. National S&T management information system public service platform. Project contains: “fishery” OR “aquatic products” OR “marine fishery” OR “aquaculture technology”; Year = 2010-2024; Project type = national key R&D, NSFC, “948” project, agricultural S&T key project. National S&T management information system public service platform. Project contains: “fishery” OR “aquatic products” OR “marine fishery” OR “aquaculture technology”; Year = 2010-2024; Project type = national key R&D, NSFC, “948” project, agricultural S&T key project. Deduplication rules Records retrieved from multiple databases were deduplicated using a three-step procedure:Step 0—CNKI-Wanfang overlap (DOI match). 30,265 overlapping records were identified by DOI match and removed as a batch.Step 1—Unique identifier matching. Records sharing the same DOI (for publications), patent number (for patents), or project approval number (for NSFC projects) were merged. This step removed 12,452 records (including 7081 within-CNKI publication duplicates, 3,402 within-Derwent patent duplicates, and 2017 within-NSFC project duplicates).Step 2—Cross-database string similarity matching. Normalized Levenshtein similarity ≥ 0.90 on title fields after lowercasing and stop-word removal. An additional 1244 cross-database duplicates were removed.Total removed across all three steps: 43,961; remaining: 65,445 unique records. Deduplication rules Records retrieved from multiple databases were deduplicated using a three-step procedure:Step 0—CNKI-Wanfang overlap (DOI match). 30,265 overlapping records were identified by DOI match and removed as a batch.Step 1—Unique identifier matching. Records sharing the same DOI (for publications), patent number (for patents), or project approval number (for NSFC projects) were merged. This step removed 12,452 records (including 7081 within-CNKI publication duplicates, 3,402 within-Derwent patent duplicates, and 2017 within-NSFC project duplicates).Step 2—Cross-database string similarity matching. Normalized Levenshtein similarity ≥ 0.90 on title fields after lowercasing and stop-word removal. An additional 1244 cross-database duplicates were removed.Total removed across all three steps: 43,961; remaining: 65,445 unique records. Step 0—CNKI-Wanfang overlap (DOI match). 30,265 overlapping records were identified by DOI match and removed as a batch. Step 0—CNKI-Wanfang overlap (DOI match). 30,265 overlapping records were identified by DOI match and removed as a batch. Step 1—Unique identifier matching. Records sharing the same DOI (for publications), patent number (for patents), or project approval number (for NSFC projects) were merged. This step removed 12,452 records (including 7081 within-CNKI publication duplicates, 3,402 within-Derwent patent duplicates, and 2017 within-NSFC project duplicates). Step 1—Unique identifier matching. Records sharing the same DOI (for publications), patent number (for patents), or project approval number (for NSFC projects) were merged. This step removed 12,452 records (including 7081 within-CNKI publication duplicates, 3,402 within-Derwent patent duplicates, and 2017 within-NSFC project duplicates). Step 2—Cross-database string similarity matching. Normalized Levenshtein similarity ≥ 0.90 on title fields after lowercasing and stop-word removal. An additional 1244 cross-database duplicates were removed. Step 2—Cross-database string similarity matching. Normalized Levenshtein similarity ≥ 0.90 on title fields after lowercasing and stop-word removal. An additional 1244 cross-database duplicates were removed. Total removed across all three steps: 43,961; remaining: 65,445 unique records. Total removed across all three steps: 43,961; remaining: 65,445 unique records. Exclusion criteria (E1–E4)E1 off-topic: not directly about fishery S&T innovation (e.g., fishery economics, policy, sociology, pure biology without tech content). Assessed at both screening (title/abstract) and eligibility (full-text) phases.E2 incomplete metadata: missing title, abstract, or keywords. Assessed at both screening (title/abstract) and eligibility (full-text) phases.E3 non-invention patent: utility models and design patents excluded; only invention patents (IPC code A01K) retained. Assessed at eligibility (full-text) phase.E4 out-of-time: year outside the 2010–2024 window. No records were excluded on this criterion, as the search was constrained to 2010-2024. Exclusion criteria (E1–E4)E1 off-topic: not directly about fishery S&T innovation (e.g., fishery economics, policy, sociology, pure biology without tech content). Assessed at both screening (title/abstract) and eligibility (full-text) phases.E2 incomplete metadata: missing title, abstract, or keywords. Assessed at both screening (title/abstract) and eligibility (full-text) phases.E3 non-invention patent: utility models and design patents excluded; only invention patents (IPC code A01K) retained. Assessed at eligibility (full-text) phase.E4 out-of-time: year outside the 2010–2024 window. No records were excluded on this criterion, as the search was constrained to 2010-2024. E1 off-topic: not directly about fishery S&T innovation (e.g., fishery economics, policy, sociology, pure biology without tech content). Assessed at both screening (title/abstract) and eligibility (full-text) phases. E1 off-topic: not directly about fishery S&T innovation (e.g., fishery economics, policy, sociology, pure biology without tech content). Assessed at both screening (title/abstract) and eligibility (full-text) phases. E2 incomplete metadata: missing title, abstract, or keywords. Assessed at both screening (title/abstract) and eligibility (full-text) phases. E2 incomplete metadata: missing title, abstract, or keywords. Assessed at both screening (title/abstract) and eligibility (full-text) phases. E3 non-invention patent: utility models and design patents excluded; only invention patents (IPC code A01K) retained. Assessed at eligibility (full-text) phase. E3 non-invention patent: utility models and design patents excluded; only invention patents (IPC code A01K) retained. Assessed at eligibility (full-text) phase. E4 out-of-time: year outside the 2010–2024 window. No records were excluded on this criterion, as the search was constrained to 2010-2024. E4 out-of-time: year outside the 2010–2024 window. No records were excluded on this criterion, as the search was constrained to 2010-2024. Cohen’s κ scoring process Two trained RAs independently screened all 65,445 post-deduplication records. Cohen’s κ: Discrepancies (11.4% of records, corresponding to 7,461 of the 65,445 post-deduplication records) were resolved through consultation with a senior domain expert (co-author, 15+ years in fishery S&T research). The arbitration procedure followed a consensus-based protocol: Each RA recorded their independent screening decision and confidence level (high/medium/low) for each discrepant record.Records with high-confidence agreement were retained as-is.Records with disagreement were discussed in a joint session; if consensus was reached, the agreed decision was recorded.If no consensus was reached after discussion (7.2% of discrepant records), the senior expert’s decision prevailed. Post-arbitration inter-rater reliability was reassessed, yielding κ = 0.91, indicating near-perfect agreement. Cohen’s κ scoring process Two trained RAs independently screened all 65,445 post-deduplication records. Cohen’s κ: Discrepancies (11.4% of records, corresponding to 7,461 of the 65,445 post-deduplication records) were resolved through consultation with a senior domain expert (co-author, 15+ years in fishery S&T research). The arbitration procedure followed a consensus-based protocol: Each RA recorded their independent screening decision and confidence level (high/medium/low) for each discrepant record.Records with high-confidence agreement were retained as-is.Records with disagreement were discussed in a joint session; if consensus was reached, the agreed decision was recorded.If no consensus was reached after discussion (7.2% of discrepant records), the senior expert’s decision prevailed. Post-arbitration inter-rater reliability was reassessed, yielding κ = 0.91, indicating near-perfect agreement. Each RA recorded their independent screening decision and confidence level (high/medium/low) for each discrepant record. Each RA recorded their independent screening decision and confidence level (high/medium/low) for each discrepant record. Records with high-confidence agreement were retained as-is. Records with high-confidence agreement were retained as-is. Records with disagreement were discussed in a joint session; if consensus was reached, the agreed decision was recorded. Records with disagreement were discussed in a joint session; if consensus was reached, the agreed decision was recorded. If no consensus was reached after discussion (7.2% of discrepant records), the senior expert’s decision prevailed. If no consensus was reached after discussion (7.2% of discrepant records), the senior expert’s decision prevailed. Named entity recognition and entity normalization Four entity types were defined for fishery S&T domain NER (Table 2): Table 2Four entity types for S&T domain.Entity typeTagDefinitionExampleTechnology keywordTKStandardized technical terms describing methods, tools, or processes“Recirculating aquaculture system”Project namePNTitles of funded research projectsNational Key R&D Program on Marine RanchingInstitutionINSUniversities, research institutes, enterprises“Chinese Academy of Fishery Sciences”ResearcherPERAuthors, inventors, principal investigatorsNames extracted from publication metadata Four entity types for S&T domain. Annotation and model training The BERT-Base-Chinese-NER model40 was fine-tuned on a domain-specific corpus constructed as follows: Annotation corpus. Two domain experts manually annotated 5,000 sentences randomly sampled from the corpus, following the BIO tagging scheme. Inter-annotator agreement, measured by Cohen’s kappa, was κ = 0.89, indicating substantial agreement. Discrepancies were resolved through consensus discussion. Annotation corpus. Two domain experts manually annotated 5,000 sentences randomly sampled from the corpus, following the BIO tagging scheme. Inter-annotator agreement, measured by Cohen’s kappa, was κ = 0.89, indicating substantial agreement. Discrepancies were resolved through consensus discussion. Fine-tuning procedure. Pre-trained BERT-Base-Chinese weights were fine-tuned using the following hyperparameters: learning rate = 2e-5, batch size = 32, epochs = 10, early stopping on validation F1. The model was validated on a held-out 20% of the annotated corpus. Fine-tuning procedure. Pre-trained BERT-Base-Chinese weights were fine-tuned using the following hyperparameters: learning rate = 2e-5, batch size = 32, epochs = 10, early stopping on validation F1. The model was validated on a held-out 20% of the annotated corpus. Performance metrics. On the held-out test set (Table 3): Performance metrics. On the held-out test set (Table 3): Table 3Performance of BERT-Base-Chinese NER.Entity typePrecisionRecallF1-scoreTechnology keyword (TK)0.9120.8870.899Project name (PN)0.9450.9380.941Institution (INS)0.9670.9540.960Researcher (PER)0.9780.9710.975Overall0.9510.9370.944 Performance of BERT-Base-Chinese NER. Cross-temporal validation To assess NER performance stability across the 2010–2024 temporal range, we conducted a stratified temporal validation by training on data from 2010–2019 and testing on 2020-2024 separately. The overall F1-score remained above 0.90 across all test periods, with a maximum decline of 3.2% in TK entity recognition for emerging technology terms (e.g., “digital twin,” “large language model”) not present in the training vocabulary. Entity normalization Entity normalization addressed two challenges: (1) Name variants (e.g., “Chinese Academy of Fishery Sciences” vs. “CAFS”) were aligned using a predefined thesaurus supplemented by string matching and cosine similarity in sentence embedding space (using sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2) with a similarity threshold of 0.85. (2) Technology keyword normalization merged synonyms (e.g., “IoT” = “Internet of Things”) and handled hypernymy/hyponymy relations using a fishery-domain ontology constructed from CNKI subject headings and the AGROVOC thesaurus. Based on manual inspection of a random sample of 500 entities, the upper bound of normalization errors is estimated at approximately 4.3%, primarily from ambiguous institutional abbreviations and multi-sense technology terms. Data were temporally sliced at the year level: for a publication from 2020, all authors, institutions, keywords, and their relationships are assigned to the 2020 graph snapshot. Negative sample generation In the link prediction task, constructing negative samples (non-existent edges) is critical for training a well-specified classifier. Following best practices in link prediction literature41, for each positive-sample edge , we generate negative samples as follows: Negative sample strategy. For each positive-sample edge , five candidate node pairs are randomly selected from all node pairs that do not co-occur in the same time-step graph snapshot , ensuring temporal comparability. The 5-candidate filter reduces the probability of accidentally sampling weak-positive pairs (two technologies that co-occurred rarely but genuinely) as negative samples. From these five candidates, one negative edge is randomly retained to maintain a positive-to-negative ratio of 1:1. Importantly, no model-predicted probabilities are computed during this sampling step, and the retained negative is not selected by hardness; this procedure is therefore a 5-candidate filtered random negative sampling strategy rather than online hard negative mining. We verified that this strategy yields AUC = 0.934 ± 0.002, compared to AUC = 0.931 ± 0.003 for direct 1:1 sampling (a consistent +0.003 AUC advantage across 5 seeds). We note that this random selection from the 5-candidate pool differs from selecting the lowest-PMI candidate (which yields AUC = 0.932 ± 0.003, statistically indistinguishable from random selection), confirming that the 5-candidate advantage stems from the filtering effect rather than from a specific candidate selection criterion within the pool. Negative sample strategy. For each positive-sample edge , five candidate node pairs are randomly selected from all node pairs that do not co-occur in the same time-step graph snapshot , ensuring temporal comparability. The 5-candidate filter reduces the probability of accidentally sampling weak-positive pairs (two technologies that co-occurred rarely but genuinely) as negative samples. From these five candidates, one negative edge is randomly retained to maintain a positive-to-negative ratio of 1:1. Importantly, no model-predicted probabilities are computed during this sampling step, and the retained negative is not selected by hardness; this procedure is therefore a 5-candidate filtered random negative sampling strategy rather than online hard negative mining. We verified that this strategy yields AUC = 0.934 ± 0.002, compared to AUC = 0.931 ± 0.003 for direct 1:1 sampling (a consistent +0.003 AUC advantage across 5 seeds). We note that this random selection from the 5-candidate pool differs from selecting the lowest-PMI candidate (which yields AUC = 0.932 ± 0.003, statistically indistinguishable from random selection), confirming that the 5-candidate advantage stems from the filtering effect rather than from a specific candidate selection criterion within the pool. Random seed specification. All random processes (negative sampling, train/val/test splitting) use a fixed random seed of 42 (torch.manual_seed, numpy.random.seed, random.seed all set to 42) for reproducibility. We verified that varying the random seed across does not change the ranking of model performance, with a maximum AUC variation of ±0.003. Random seed specification. All random processes (negative sampling, train/val/test splitting) use a fixed random seed of 42 (torch.manual_seed, numpy.random.seed, random.seed all set to 42) for reproducibility. We verified that varying the random seed across does not change the ranking of model performance, with a maximum AUC variation of ±0.003. False negative audit A critical assumption in link prediction is that non-existent edges in the current snapshot are genuinely negative (will not co-occur in the future). If this assumption is violated-i.e., if some “negative” edges are actually “future-positive” edges that will form in a later time step-then the model’s AUC may be biased in either direction. Such samples may penalize a model that correctly anticipates future links, or may otherwise distort the negative class composition. To assess the severity of this potential bias, we conducted a false negative audit.Procedure. We randomly sampled 200 “negative edges” (u, v) from the 2023 validation set (i.e., technology pairs that did not co-occur in any 2010-2023 document) and manually inspected whether u and v actually co-occur in the 2024 test set, which would indicate that they were false negatives in 2023.Results. Among the 200 sampled negative edges from 2023:12 edges (6.0%) actually co-occurred in 2024, confirming that they were false negatives in 2023. These 12 edges had mean PMI = 0.42 (vs. mean PMI = 0.08 for truly negative edges), suggesting that high-PMI non-edges are disproportionately likely to be false negatives.8 edges (4.0%) had PMI > 0 but did not co-occur in the same document in 2024, indicating “weak positive” signals that were missed by the binary negative sampling strategy.Impact on AUC estimation. We no longer interpret these future-positive negatives as necessarily inflating AUC. The direction of the effect is ambiguous: future-positive pairs are often harder cases (as suggested by their higher mean PMI) and may reduce the measured AUC of a model that assigns them high probabilities, even though such probabilities may be desirable for forecasting. We therefore treat the 6.0% future-positive rate as label noise in the validation negatives rather than as a quantified inflation term. The reported AUC should be interpreted with this limitation in mind, and a fully corrected estimate would require rebuilding the validation set after removing or relabeling future-positive negatives.Mitigation strategy. Future work could incorporate dynamic negative sampling, where negative edges are periodically re-evaluated against new data and promoted to positive edges if they later co-occur, thereby reducing the false negative rate over time. False negative audit A critical assumption in link prediction is that non-existent edges in the current snapshot are genuinely negative (will not co-occur in the future). If this assumption is violated-i.e., if some “negative” edges are actually “future-positive” edges that will form in a later time step-then the model’s AUC may be biased in either direction. Such samples may penalize a model that correctly anticipates future links, or may otherwise distort the negative class composition. To assess the severity of this potential bias, we conducted a false negative audit.Procedure. We randomly sampled 200 “negative edges” (u, v) from the 2023 validation set (i.e., technology pairs that did not co-occur in any 2010-2023 document) and manually inspected whether u and v actually co-occur in the 2024 test set, which would indicate that they were false negatives in 2023.Results. Among the 200 sampled negative edges from 2023:12 edges (6.0%) actually co-occurred in 2024, confirming that they were false negatives in 2023. These 12 edges had mean PMI = 0.42 (vs. mean PMI = 0.08 for truly negative edges), suggesting that high-PMI non-edges are disproportionately likely to be false negatives.8 edges (4.0%) had PMI > 0 but did not co-occur in the same document in 2024, indicating “weak positive” signals that were missed by the binary negative sampling strategy.Impact on AUC estimation. We no longer interpret these future-positive negatives as necessarily inflating AUC. The direction of the effect is ambiguous: future-positive pairs are often harder cases (as suggested by their higher mean PMI) and may reduce the measured AUC of a model that assigns them high probabilities, even though such probabilities may be desirable for forecasting. We therefore treat the 6.0% future-positive rate as label noise in the validation negatives rather than as a quantified inflation term. The reported AUC should be interpreted with this limitation in mind, and a fully corrected estimate would require rebuilding the validation set after removing or relabeling future-positive negatives.Mitigation strategy. Future work could incorporate dynamic negative sampling, where negative edges are periodically re-evaluated against new data and promoted to positive edges if they later co-occur, thereby reducing the false negative rate over time. Procedure. We randomly sampled 200 “negative edges” (u, v) from the 2023 validation set (i.e., technology pairs that did not co-occur in any 2010-2023 document) and manually inspected whether u and v actually co-occur in the 2024 test set, which would indicate that they were false negatives in 2023. Procedure. We randomly sampled 200 “negative edges” (u, v) from the 2023 validation set (i.e., technology pairs that did not co-occur in any 2010-2023 document) and manually inspected whether u and v actually co-occur in the 2024 test set, which would indicate that they were false negatives in 2023. Results. Among the 200 sampled negative edges from 2023:12 edges (6.0%) actually co-occurred in 2024, confirming that they were false negatives in 2023. These 12 edges had mean PMI = 0.42 (vs. mean PMI = 0.08 for truly negative edges), suggesting that high-PMI non-edges are disproportionately likely to be false negatives.8 edges (4.0%) had PMI > 0 but did not co-occur in the same document in 2024, indicating “weak positive” signals that were missed by the binary negative sampling strategy. Results. Among the 200 sampled negative edges from 2023:12 edges (6.0%) actually co-occurred in 2024, confirming that they were false negatives in 2023. These 12 edges had mean PMI = 0.42 (vs. mean PMI = 0.08 for truly negative edges), suggesting that high-PMI non-edges are disproportionately likely to be false negatives.8 edges (4.0%) had PMI > 0 but did not co-occur in the same document in 2024, indicating “weak positive” signals that were missed by the binary negative sampling strategy. 12 edges (6.0%) actually co-occurred in 2024, confirming that they were false negatives in 2023. These 12 edges had mean PMI = 0.42 (vs. mean PMI = 0.08 for truly negative edges), suggesting that high-PMI non-edges are disproportionately likely to be false negatives. 12 edges (6.0%) actually co-occurred in 2024, confirming that they were false negatives in 2023. These 12 edges had mean PMI = 0.42 (vs. mean PMI = 0.08 for truly negative edges), suggesting that high-PMI non-edges are disproportionately likely to be false negatives. 8 edges (4.0%) had PMI > 0 but did not co-occur in the same document in 2024, indicating “weak positive” signals that were missed by the binary negative sampling strategy. 8 edges (4.0%) had PMI > 0 but did not co-occur in the same document in 2024, indicating “weak positive” signals that were missed by the binary negative sampling strategy. Impact on AUC estimation. We no longer interpret these future-positive negatives as necessarily inflating AUC. The direction of the effect is ambiguous: future-positive pairs are often harder cases (as suggested by their higher mean PMI) and may reduce the measured AUC of a model that assigns them high probabilities, even though such probabilities may be desirable for forecasting. We therefore treat the 6.0% future-positive rate as label noise in the validation negatives rather than as a quantified inflation term. The reported AUC should be interpreted with this limitation in mind, and a fully corrected estimate would require rebuilding the validation set after removing or relabeling future-positive negatives. Impact on AUC estimation. We no longer interpret these future-positive negatives as necessarily inflating AUC. The direction of the effect is ambiguous: future-positive pairs are often harder cases (as suggested by their higher mean PMI) and may reduce the measured AUC of a model that assigns them high probabilities, even though such probabilities may be desirable for forecasting. We therefore treat the 6.0% future-positive rate as label noise in the validation negatives rather than as a quantified inflation term. The reported AUC should be interpreted with this limitation in mind, and a fully corrected estimate would require rebuilding the validation set after removing or relabeling future-positive negatives. Mitigation strategy. Future work could incorporate dynamic negative sampling, where negative edges are periodically re-evaluated against new data and promoted to positive edges if they later co-occur, thereby reducing the false negative rate over time. Mitigation strategy. Future work could incorporate dynamic negative sampling, where negative edges are periodically re-evaluated against new data and promoted to positive edges if they later co-occur, thereby reducing the false negative rate over time. Validation and test set construction. The positive-to-negative ratio in validation and test sets is maintained at 1:1. Datasets are split as: training set (2010-2022, 13 annual snapshots), validation set (2023, 1 snapshot, used for hyperparameter tuning and early stopping), and test set (2024, 1 snapshot, used for final evaluation). Validation and test set construction. The positive-to-negative ratio in validation and test sets is maintained at 1:1. Datasets are split as: training set (2010-2022, 13 annual snapshots), validation set (2023, 1 snapshot, used for hyperparameter tuning and early stopping), and test set (2024, 1 snapshot, used for final evaluation). Sensitivity analysis on negative sampling We conducted a systematic sensitivity analysis on the number of negative candidates (1, 3, 5, and 10) to justify our selection of k=5 (Fig. 4). As shown in Table 4, AUC increases monotonically from 0.921 (1-candidate) to 0.935 (10-candidate), while AP peaks at 0.929 (5-candidate) and decreases to 0.925 (10-candidate). A paired t-test comparing 5-candidate vs. 1-candidate yields t=4.32, p<0.01, confirming statistically significant improvement. Critically, increasing candidates beyond 5 not only yields negligible AUC gain (+0.001) but actually hurts AP (-0.004), suggesting an optimal trade-off at k=5 that balances performance and generalization. Sensitivity analysis on negative sampling strategy. Table 4Sensitivity analysis on negative sampling: effect of candidate number on link prediction performance. AP peaks at k=5 and decreases at k=10.Candidate numberAUCAPTraining time (s/epoch)10.9210.91512.330.9280.92414.70.9340.92918.2100.9350.92527.5Δ (5 vs. 1)+0.013+0.014+5.9Significant values are in bold. Sensitivity analysis on negative sampling: effect of candidate number on link prediction performance. AP peaks at k=5 and decreases at k=10. Significant values are in bold. The computational overhead is acceptable: training time increases from 12.3s/epoch (1-candidate) to 18.2s/epoch (5-candidate, +48% overhead), while 10-candidate incurs 27.5s/epoch (+124% overhead) with degraded AP. We selected 5 as the optimal setting following the principle of diminishing marginal returns, consistent with negative sampling practices in knowledge graph embedding42,43. Hard negative mining analysis To assess whether the 5-candidate strategy merely selects less informative negative samples (which could bias evaluation without improving genuine predictive power), we conducted a hard negative mining experiment. Procedure For each positive edge , instead of using the default 5-candidate filtered random sampling procedure, we selected the single negative candidate (u, v') with the highest PMI (Pointwise Mutual Information) value among all non-neighbor pairs. This “hard negative” represents a technology pair that nearly co-occurs but does not form an edge in the current snapshot-a more challenging negative sample that tests whether the model can distinguish genuinely absent connections from near-miss co-occurrences. Results Training DTH-GNN with hard negatives yields AUC = 0.928 ± 0.004 and AP = 0.923 ± 0.003, which is lower than the 5-candidate strategy (AUC = 0.934, AP = 0.929) but higher than 1-candidate random sampling (AUC = 0.921, AP = 0.915). This pattern suggests that: The 5-candidate strategy provides a regularization effect: by sampling from a broader candidate pool, it prevents the model from overfitting to hard-negative patterns that may reflect near-miss co-occurrences rather than genuinely absent connections; The 5-candidate strategy provides a regularization effect: by sampling from a broader candidate pool, it prevents the model from overfitting to hard-negative patterns that may reflect near-miss co-occurrences rather than genuinely absent connections; Hard negatives are informative (AUC = 0.928 vs. 0.921 for easy negatives), confirming that the model learns meaningful discriminative features rather than relying on trivial co-occurrence statistics; Hard negatives are informative (AUC = 0.928 vs. 0.921 for easy negatives), confirming that the model learns meaningful discriminative features rather than relying on trivial co-occurrence statistics; However, relying exclusively on hard negatives slightly hurts performance (AUC = 0.928 vs. 0.934), suggesting that hard negatives may introduce a subtle bias toward recognizing only “near-miss” patterns. However, relying exclusively on hard negatives slightly hurts performance (AUC = 0.928 vs. 0.934), suggesting that hard negatives may introduce a subtle bias toward recognizing only “near-miss” patterns. We therefore retain the 5-candidate strategy as a balanced approach that incorporates both easy and moderately hard negatives, and recommend future work to explore adaptive negative sampling strategies that dynamically adjust difficulty based on training progress. Construction of the dynamic heterogeneous graph Graph structure definition The dynamic heterogeneous graph constructed in this study, , contains the following nodes and edges at each time step t (year). Node and edge definitions are presented in Tables 3 and 4, respectively (Table 5). Table 5Definition of nodes in the dynamic heterogeneous graph.Node typeNode definitionDynamic attributesTechnology node (T)Standardized technology keywords extracted from publications, patents, and project abstracts (e.g., “gene editing”, “recirculating aquaculture system”, “IoT”, “microalgae biodiesel”)Node centrality (degree centrality, PageRank), annual term frequency, year of first appearanceProject node (P)National- and provincial/ministerial-level research projectsFunding amount, project duration, number of participating institutionsInstitution node (I)Universities, research institutes, and enterprisesAnnual publication/patent/project count, network centralityResearcher node (R)Publication authors, patent inventors, and principal investigatorsAnnual output count, h-index, collaboration network centralityLiterature node (A)Graph-instantiated publications and patents; research projects are represented separately as project nodes (P)Patent information, publication information, authorship relations Definition of nodes in the dynamic heterogeneous graph. Edge weight calculation Edge weights are computed separately by relation type to preserve the semantics of heterogeneous links. Collaboration edges are weighted by the annual number of co-authored publications, co-filed patents, or jointly participated projects between two researchers. Affiliation, undertaking, funding, and citation edges are binary within each annual snapshot (weight = 1 if the relation is observed, 0 otherwise). Technology-technology co-occurrence edges are weighted by annual co-occurrence frequency adjusted by Pointwise Mutual Information (PMI) to reflect semantic association strength44: where p(x) and p(y) denote the probabilities of keywords x and y in the network, respectively, and p(x, y) denotes the probability of their co-occurrence in the same document. A positive PMI value indicates a positive correlation between two nodes; a negative value indicates negative correlation. To mitigate noise from low-frequency co-occurrences, PMI is set to 0 when co-occurrence frequency is three or fewer. For technology association edges, we distinguish direct text-derived edges from projected edges. A publication, patent, or project node is directly connected to a technology node when the normalized keyword appears in its title, abstract, or keyword field; the edge weight is the keyword’s TF-IDF value in that record. Institution-technology and researcher-technology edges are not literal “containment” edges. They are projected from document/project-level relations: an institution is associated with a technology through its affiliated publications, patents, or undertaken projects that contain the keyword, and a researcher is associated with a technology through authored publications, invented patents, or funded projects that contain the keyword. The projected edge weight is the sum of TF-IDF values over linked records within the same year, normalized by the number of linked records for that institution or researcher. All edge weights are computed independently within each annual snapshot before being passed to the heterogeneous graph encoder. Graph construction results After data integration and graph construction, 15 annual graph snapshots spanning 2010–2024 were produced. Table 7 summarizes the statistical characteristics of annual heterogeneous graphs. To avoid ambiguity between corpus-level records and graph-level nodes, we distinguish the two accounting units explicitly. The final source corpus contains 43,659 records (32,847 publications, 8956 patents, and 1856 projects), whereas Table 7 reports the graph-instantiated nodes in each annual snapshot after entity extraction, normalization, and edge construction. In the 2024 snapshot, the graph contains 33,862 literature nodes (publications and patents instantiated as A-type nodes) and 1431 project nodes (P-type nodes), for 35,293 document/project nodes. The remaining 47,710 nodes are extracted entities: 29,876 technology nodes, 2156 institution nodes, and 15,678 researcher nodes. Thus, the total graph-node count (83,003) exceeds the source-record count because technology, institution, and researcher entities are modeled as additional nodes. Conversely, the literature-node count alone is not expected to equal the total source-record count because projects are counted separately and only records instantiated with valid normalized graph relations enter the annual graph topology as document/project nodes. As shown in Table 7, the number of technology nodes grew from 1,245 in 2010 to 29,876 in 2024 (CAGR: 25.5%; average annual growth: 26.0%). To better characterize the connectivity dynamics, we compute the edge-to-node ratio (ENR, |E|/|V|): 7.62 in 2010, 8.84 in 2015, 13.56 in 2020, and 13.64 in 2024 (see Table 7 for complete annual values). The ENR increased by 79% over this period (7.62 → 13.64), indicating that the network becomes structurally denser: per-node connectivity grew faster than the number of nodes. While network density (proportion of possible edges) decreases due to the combinatorial explosion of possible edges as |V| grows, the average per-node connectivity increases substantially, indicating that the network becomes structurally denser in practice. The new technology nodes (emerging technologies) form sparse individual connections but contribute substantially to the network’s edge volume through heterogeneous mediating connections (institution, project, researcher, and literature nodes), validating the value of DTH-GNN’s heterogeneous encoding (Fig. 5). Edge-to-node ratio over time (2010–2024). Technology-technology co-occurrence edges account for approximately 65% of all edges, constituting the primary structural characteristic of the network. Analysis of 2024 graph structure reveals that the degree distribution of technology nodes follows a typical power-law pattern (consistent with the Barabási-Albert preferential attachment model45), indicating that a small number of core technology nodes (e.g., “gene editing”, “IoT”, “artificial intelligence”) are connected to a large number of other technologies. The network also exhibits small-world characteristics (average shortest path length = 3.2, clustering coefficient = 0.34, compared to random graph baselines of 4.8 and 0.02 respectively), consistent with the general properties of innovation networks documented in prior literature4,20. PageRank analysis of institution nodes shows that the Chinese Academy of Fishery Sciences, Ocean University of China, and Qingdao Agricultural University occupy central positions in the network. Mutual information analysis—feature definitions and estimation procedure To quantify the relative association between temporal, structural, and attribute features and link existence, we conducted a mutual information (MI) analysis following the procedure of Faivishevsky and Goldberger46. We treat this analysis as descriptive rather than causal because the feature groups differ in dimensionality and information content. Feature definitions: Three non-overlapping feature categories are defined: Temporal features (): node age (, years since first appearance), edge recency (, years since most recent co-occurrence), and temporal clustering coefficient (, proportion of node v’s neighbors that co-occur with each other within a 2-year window). Temporal features (): node age (, years since first appearance), edge recency (, years since most recent co-occurrence), and temporal clustering coefficient (, proportion of node v’s neighbors that co-occur with each other within a 2-year window). Structural features (): degree centrality, PageRank, betweenness centrality, eigenvector centrality, Katz index, and Adamic-Adar index. All computed on the final graph snapshot (i.e., static features, not time-varying). Structural features (): degree centrality, PageRank, betweenness centrality, eigenvector centrality, Katz index, and Adamic-Adar index. All computed on the final graph snapshot (i.e., static features, not time-varying). Attribute features (): BERT keyword embeddings from Eq. (6), project funding level (one-hot, 3 levels), institution type (one-hot, 5 categories). Before MI estimation, the 256-dimensional BERT embeddings were standardized and reduced by principal component analysis (PCA) to 32 components to reduce nearest-neighbor distance instability in high dimensions; categorical one-hot variables were then concatenated to the reduced representation. Attribute features (): BERT keyword embeddings from Eq. (6), project funding level (one-hot, 3 levels), institution type (one-hot, 5 categories). Before MI estimation, the 256-dimensional BERT embeddings were standardized and reduced by principal component analysis (PCA) to 32 components to reduce nearest-neighbor distance instability in high dimensions; categorical one-hot variables were then concatenated to the reduced representation. Preprocessing and MI estimator: all continuous variables were z-score standardized within the corresponding annual snapshot before MI estimation. We used the Kraskov–Stögbauer-Grassberger (KSG) estimator47 with k=5 nearest neighbors. For this estimator, k controls the bias-variance trade-off: larger k reduces variance but increases bias when the sample size is limited. Given the scale of our link-prediction sample, k=5 was selected as a conservative choice consistent with prior work on MI estimation for link prediction tasks46. We additionally verified that varying does not materially change the qualitative ordering of feature-category associations. Procedure: (a) For each feature category , compute the MI where is the binary link existence variable at the test time step. (b) Normalize by total MI to obtain percentage shares. (c) Report 95% bootstrap confidence intervals (B=200 resamples). (d) Conduct robustness checks by varying the BERT PCA dimensionality (16, 32, and 64 components) and by regrouping features into four categories (temporal, graph topology, link heuristics, and textual attributes). Results and interpretation: Temporal features account for 31.8% of total MI (95% CI: [28.4%, 35.1%]), structural features 29.1% [26.2%, 32.0%], and attribute features 19.7% [17.1%, 22.3%]. The remaining 19.4% [16.8%, 21.9%] is reported as a non-additive residual caused by overlap among feature groups and estimator limitations; it should not be interpreted as a directly measured contribution of feature interactions or model nonlinearity. The temporal-feature estimate is slightly higher than the structural-feature estimate, but their confidence intervals overlap. We therefore interpret the MI analysis as evidence that temporal features carry substantial discriminative information, not as a formal decomposition proving that temporal features are uniquely or causally dominant. DTH-GNN model To effectively learn from the dynamic heterogeneous graph described above, this study develops a Documents-based Temporal Heterogeneous Graph Neural Network (DTH-GNN) model. The model comprises two main components: a heterogeneous graph encoder and a temporal evolution module. Heterogeneous graph encoder At each time step t, an R-GCN-based encoder is used to learn node representations. For any node v in graph , the representation at layer l+1 is updated as Eq. (5): where is the set of all relation types, is the set of neighbors of node v under relation r, is the learnable weight matrix specific to relation r, is the self-transformation weight matrix, and is a normalization constant. Through multi-layer stacking, each node aggregates rich information from different types of neighbors, generating node embeddings that encode heterogeneous structural information. Node feature initialization Each node type receives a distinct initialization strategy reflecting its semantic properties: Technology nodes (T): feature vectors are initialized as 256-dimensional embeddings derived from the averaged BERT embeddings of the technology keyword’s contextual occurrences in the corpus. Specifically, for technology keyword v with n occurrences across publications/patents/projects: where is the textual context (title + abstract) of the j-th occurrence, and denotes the [CLS] token output from BERT-Base-Chinese. Project nodes (P): initialized as 256-dimensional vectors from the BERT embedding of the project title + abstract, augmented with one-hot encoded categorical features (funding level: national/provincial/ministerial; research field: 7 categories per Table 6). Institution nodes (I): initialized as 256-dimensional vectors from the BERT embedding of the institution name, concatenated with a 16-dimensional structural feature vector (annual publication count, patent count, project count, PageRank centrality from the previous snapshot). Researcher nodes (R): initialized as 256-dimensional vectors from the BERT embedding of the researcher’s most frequent publication title, concatenated with a 16-dimensional feature vector (annual output count, h-index, collaboration degree). Literature nodes (A): initialized as 256-dimensional vectors from the BERT embedding of the document title + abstract. All feature vectors are projected to the 256-dimensional hidden space via a learnable linear transformation before being fed into the R-GCN encoder. Temporal evolution module As fishery S&T innovation networks evolve dynamically over time-with new nodes emerging, existing nodes disappearing, and edge weights changing-the node embeddings encoded by R-GCN at different time steps may suffer from semantic space drift. Direct comparison or propagation of these embeddings across time steps would lead to distorted temporal information and inconsistent evolutionary patterns. Therefore, an explicit mechanism is required to ensure that and reside in the same embedding space and maintain semantic comparability. To capture the temporal evolution of node representations, the embeddings of each node v at different time steps, , are treated as a time series. Gated Recurrent Units (GRU)34 are employed to model this sequence. At time step t, the GRU update is Eqs. (7)–(10): where is the current-time-step node embedding from the heterogeneous graph encoder, and is the hidden state of GRU at the previous time step. The final output is the ultimate dynamic embedding of node v at time step t, jointly encoding current graph structural information and historical evolution information. The GRU ensures embedding space alignment through its gating mechanisms: the Update Gate controls the retention ratio of historical information , while the Reset Gate determines the utilization degree of current input . Three key design elements guarantee cross-temporal consistency: Shared GRU parameters: all time steps employ the same set of transformation weights , ensuring consistent embedding transformations across the temporal sequence. Recurrent dependency structure: the final embedding explicitly depends on through the chain rule: This creates a chained constraint that forces embeddings at all time steps to maintain fixed dimensionality (256-dimensional in our implementation) and semantic comparability. End-to-end training: the R-GCN encoder and GRU module are jointly optimized, where backpropagation through time (BPTT) enforces temporal consistency by minimizing: For newly emerging nodes without historical states, is initialized as a zero vector. We conducted a sensitivity analysis to assess the impact of this initialization strategy on model performance and training dynamics (Table 8). Table 8Alternative initialization strategies tested.InitializationAUCAPΔ vs. ZeroZero vector0.9340.928–Xavier initialization0.9320.925-0.002Mean of existing embeddings0.9290.922-0.005Learned initialization0.9360.930+0.002 Alternative initialization strategies tested. The zero initialization performs competitively with learned initialization, with a marginal advantage for the learned approach (+0.002 AUC). Given the minimal performance gain and the added implementation complexity, we retained zero initialization as the default. Notably, newly emerging nodes represent only 12.3% of nodes in the test set, limiting the maximum possible impact of initialization strategy. Hot-spot pathway identification via link prediction (1) From edge predictions to pathways: a three-layer empirical justification The pathway identification problem presents a methodological challenge: DTH-GNN’s high edge-level prediction accuracy (AUC = 0.934) does not, a priori, guarantee that sequences of predicted edges form semantically coherent technological trajectories. The pathway construction problem is fundamentally combinatorial: even with perfect edge-level predictions, the number of possible paths grows exponentially with length, and not all high-probability edge sequences correspond to meaningful technological trajectories. We address this challenge through a three-layer empirical justification that combines algorithmic analysis, empirical validation, and domain expert evaluation. We note that none of these layers constitute formal mathematical guarantees; rather, they provide converging empirical evidence that the greedy pathway construction produces meaningful results on our specific graph structure: Layer 1: empirical validation of pathway stability. We establish that the greedy pathway construction does not accumulate prediction errors. Validation 1 empirically demonstrates that under the calibrated confidence threshold (τ = 0.50), each edge extension preserves rather than degrades pathway quality, measured by the geometric mean score S(P). This property, which we term monotonic pathway coherence, ensures that the greedy algorithm’s sequential construction does not compound edge-level prediction errors into pathway-level degradation. Layer 2: empirical validation of approximation quality. We establish that the greedy algorithm’s outputs are robust to algorithmic variation. Validation 2 empirically confirms that under the observed probability separation conditions. the greedy-optimal pathway converges with beam search (beam width k=3) for top-ranked candidates. This greedy-beam equivalence confirms that the greedy assumption does not sacrifice solution quality for computational efficiency. Layer 3: domain validation. Theoretically grounded pathways are validated by domain experts using a modified Delphi method. Expert evaluation confirms high pathway coherence scores (Technical Feasibility: 5.87 ± 0.64; Application Value: 6.13 ± 0.35) and substantial inter-rater agreement (Krippendorff’s α = 0.804), providing external validation that the formally constructed pathways correspond to meaningful technological trajectories recognized by domain practitioners. The residual gap between formal guarantees and empirical pathway quality reflects the inherent limitation of treating technological trajectories as optimization targets: the algorithm optimizes a proxy objective (geometric mean edge probability) rather than “meaningfulness” per se. This limitation is acknowledged and addressed through dual validation (theoretical + expert). (2) Hypothesis-component correspondence The two theoretical hypotheses proposed in the Theoretical Hypotheses Section provide intuition rather than formal guarantees for the pathway construction algorithm: H1 (Heterogeneity Value-Added Hypothesis) motivates the use of R-GCN for heterogeneous graph encoding, ensuring that predicted edges are informed by multi-actor relationships; H2 (Temporal Dynamics Primacy Hypothesis) motivates the sequential reinforcement logic in Algorithm 1, reflecting the path-dependent evolution patterns captured by the GRU module. (2) Pathway construction logic: theoretical and algorithmic grounding The greedy sequential strategy in Algorithm 1 is motivated by the sequential reinforcement principle: once a technology node establishes strong connections with prior nodes in a pathway, it becomes increasingly likely to extend connections with subsequent nodes. This principle draws from three interconnected theoretical perspectives (i)Path dependence and lock-in. Dosi’s concept of technological trajectory implies that early technological choices constrain future options through increasing returns and learning curve effects22. In network terms, this suggests that nodes with high connectivity in earlier time steps will accumulate advantage, leading to sequential strengthening of pathway coherence. Path dependence and lock-in. Dosi’s concept of technological trajectory implies that early technological choices constrain future options through increasing returns and learning curve effects22. In network terms, this suggests that nodes with high connectivity in earlier time steps will accumulate advantage, leading to sequential strengthening of pathway coherence. (ii)Network effects and indirect benefits. As formalized in network economics48, the value of a technology node increases with the number and quality of its connections. This creates positive feedback loops where well-connected nodes attract additional connections, reinforcing pathway sequences. Network effects and indirect benefits. As formalized in network economics48, the value of a technology node increases with the number and quality of its connections. This creates positive feedback loops where well-connected nodes attract additional connections, reinforcing pathway sequences. (iii)Algorithmic approximation theory. Formally, the maximum pathway score problem-defined as subject to the constraints in Definition 1-is NP-hard in general (reducible from the longest path problem). The greedy algorithm achieves an (1 - 1/e)-approximation under the assumption that edge probabilities are independent and the pathway score is multiplicative. Critically, our empirical observation that greedy-optimal pathways converge with beam-3-optimal pathways (Table 16) suggests that the actual approximation ratio exceeds the worst-case bound for realistically-calibrated DTH-GNN outputs, where edge probabilities exhibit sufficient separation (mean separation: 0.18 ± 0.04) to satisfy the probability separation condition observed in Validation 2. Algorithmic approximation theory. Formally, the maximum pathway score problem-defined as subject to the constraints in Definition 1-is NP-hard in general (reducible from the longest path problem). The greedy algorithm achieves an (1 - 1/e)-approximation under the assumption that edge probabilities are independent and the pathway score is multiplicative. Critically, our empirical observation that greedy-optimal pathways converge with beam-3-optimal pathways (Table 16) suggests that the actual approximation ratio exceeds the worst-case bound for realistically-calibrated DTH-GNN outputs, where edge probabilities exhibit sufficient separation (mean separation: 0.18 ± 0.04) to satisfy the probability separation condition observed in Validation 2. The greedy algorithm is therefore characterized as an empirically validated approximation algorithm, whose practical effectiveness is supported by: (i) empirical observations that greedy-beam equivalence holds for top-3 pathways on our calibrated graph (Validation 2), and (ii) expert evaluation that confirms pathway meaningfulness. The formal (1-1/e)-approximation bound (Validation 1) applies under idealized independence assumptions; the empirical validation demonstrates that our actual graph structure satisfies conditions stronger than the worst case. Table 16Comparison of pathway construction algorithms.AlgorithmTop-3 pathways identifiedMean edge probabilityComputational timeGreedy (Algorithm 1)Identical to greedy0.7212 minBeam-3Identical to greedy0.7161 minBeam-5Top-3 stable; ranks 4–10 reordered0.70312 min Comparison of pathway construction algorithms. (3) Link prediction formulation Hot-spot pathway identification refers to systematically analyzing the evolutionary trajectories of technology or knowledge domains, identifying key technology topics, methodologies, or problem evolution paths, and predicting development trends to support R&D direction selection and strategic decision-making. Semantic analysis based on Subject-Action-Object (SAO) triples and domain-specific knowledge networks is one mainstream approach. Because hot-spot pathways are inherently contained within knowledge-network triples, this study abstracts the hot-spot pathway identification problem as a future-link path prediction task between technology nodes (T-type). Specifically, the model predicts whether a co-occurrence relation will exist between technology nodes and at time step t+1, using a bilinear interaction decoder with a learnable relation-specific matrix : where is the sigmoid function, and is initialized as an identity matrix and jointly optimized with the encoder during end-to-end training. We selected the bilinear decoder over the simpler dot product () and MLP alternatives based on a pilot comparison on the validation set (bilinear: AUC = 0.934 ± 0.002, dot product: AUC = 0.921 ± 0.004, MLP: AUC = 0.929 ± 0.003; 5 seeds each). The MLP architecture consists of two hidden layers (256 → 128 → 1) with ReLU activation and dropout (rate = 0.1). Statistical significance (paired t-test, Bonferroni-corrected): bilinear vs. MLP, p = 0.012; bilinear vs. dot product, p = 0.003. This confirms that the bilinear form provides a favorable trade-off between expressiveness (allowing asymmetric interaction patterns) and parameter efficiency ( parameters vs. for an L-layer MLP). The model is trained end-to-end using binary cross-entropy loss, maximizing predicted probabilities for true edges and minimizing them for non-edges7. To transform link prediction scores into interpretable hot-spot pathways, we propose a greedy sequential path reconstruction algorithm (Algorithm 1): Greedy sequential path reconstruction algorithm. Seed node selection rationale: the seed set comprises technology nodes whose degree exceeds one standard deviation above the mean in the most recent graph snapshot (2024). This heuristic selects high-centrality “hub” technologies that are structurally positioned to initiate meaningful pathways. In our graph, this criterion yields seed nodes from 29,876 technology nodes (9.5%). Pathway ranking stability analysis: we analyzed how pathway rankings change when the seed set is expanded from top 9.5% (degree ) to top 20% (degree ) by degree. Quantitative results are summarized in Table 9. Key findings: (i) Top-3 pathways are fully stable (100%) across all expansion thresholds, confirming greedy-beam equivalence for primary predictions; (ii) Top-5 stability ranges from 87-100% (mean 93%), with changes concentrated at aggressive expansion thresholds; (iii) Top-10 sensitivity increases monotonically (100% → 80% → 53%), as moderate-degree seed nodes capture emerging trajectories inaccessible to high-centrality hubs. Table 9Seed set expansion stability: percentage of years (2010-2024) where pathway sets remain identical under different degree thresholds.Seed thresholdTop-3 stableTop-5 stableTop-10 stableTop-9.5% ()100% (15/15)100% (15/15)100% (15/15)Top-15%100% (15/15)93% (14/15)80% (12/15)Top-20% ()100% (15/15)87% (13/15)53% (8/15) “Stable” = identical pathway set (not just overlap). Baseline uses seed set at degree (top-9.5%). Lower-ranked pathways show expected sensitivity to seed selection as moderate-degree nodes capture emerging trajectories. Seed set expansion stability: percentage of years (2010-2024) where pathway sets remain identical under different degree thresholds. “Stable” = identical pathway set (not just overlap). Baseline uses seed set at degree (top-9.5%). Lower-ranked pathways show expected sensitivity to seed selection as moderate-degree nodes capture emerging trajectories. Pathway stability under edge perturbation: to assess whether the identified hot-spot pathways are robust to edge-level prediction errors-rather than being fragile artifacts of overfitted predictions-we conducted a perturbation test by systematically adding Gaussian noise to edge probabilities. Procedure: for each of the top-10 predicted pathways, we randomly selected 10% of constituent edge probabilities and added Gaussian noise to each selected edge. We then re-ran Algorithm 1 with the perturbed probabilities (keeping all non-selected edges unchanged) and compared the resulting pathways against the original pathways. This procedure was repeated 100 times for each pathway. Metric: we report the pathway overlap ratio, defined as the fraction of shared nodes between the original and perturbed pathways: overlap = . Results: the top-3 pathways exhibit high stability: Smart Aquaculture (mean overlap = 87.5%, SD = 4.2%), Green Seed Industry (82.3%, SD = 5.1%), and Ecological Fisheries (79.8%, SD = 6.3%). Pathways ranked 4-10 show moderate stability (mean overlap = 71.4%, SD = 8.7%). No perturbation caused a top-3 pathway to be displaced from the top-3 ranking, confirming that the greedy algorithm captures robust pathway structures rather than overfitting to noisy edge predictions. Interpretation: the perturbation test provides empirical evidence that the identified pathways are stable to moderate perturbations in edge probability estimates. This result, combined with the greedy-beam equivalence (Validation 2) and expert validation, supports the conclusion that the greedy pathway construction produces reliable top-3 candidates for policy prioritization purposes. Algorithmic validity: Algorithm 1 is characterized as an empirically validated approximation algorithm, not a procedure with formal convergence guarantees. Validation 1 provides empirical evidence that the greedy construction does not systematically accumulate edge-level prediction errors on our specific graph structure (98.2% of extensions preserve monotonicity). Validation 2 provides empirical evidence of greedy-beam equivalence for top-3, confirming that the greedy solution is empirically equivalent to exhaustive beam search for top-ranked pathways on our calibrated graph. Together, these results provide empirical support-not formal mathematical proof-for the greedy algorithm’s pathway construction, complementing the a posteriori expert validation and the pathway stability analysis reported above. The residual gap between formal guarantees and empirical pathway quality reflects the inherent limitation of treating technological trajectories as optimization targets: the algorithm optimizes a proxy objective (geometric mean edge probability) rather than “meaningfulness” per se. We therefore characterize the greedy algorithm as an empirically validated heuristic whose practical effectiveness is supported by three converging lines of evidence-but whose theoretical guarantees are limited to our specific graph topology and probability calibration, and do not generalize to arbitrary graphs without re-validation. In this algorithm, confidence threshold (τ) was set to 0.5 after sensitivity analysis (Algorithm 2). Table 10 reports the performance of the pathway reconstruction algorithm under different τ values on the 2023 validation set. The optimal τ = 0.50 is selected, which yields the highest F1-score (0.847) among all candidates. Results are robust within the range [0.45, 0.55]. Table 10Sensitivity analysis of confidence threshold τ on validation set F1-score.Candidate idx.123456789Confidence threshold (τ)0.300.350.400.450.500.550.600.650.70F1-score0.7120.7890.8210.8400.8470.8380.8150.7920.756 Sensitivity analysis of confidence threshold τ on validation set F1-score. On the convergence of the F1-optimal threshold and the random-chance baseline: because the training data maintains a 1:1 positive-to-negative ratio and the model is trained with binary cross-entropy loss, the BCE gradient landscape has no systematic bias to shift the mean logit away from zero-the point at which the sigmoid output equals 0.50. Consequently, the predicted probability distribution naturally bifurcates near 0.50, making τ = 0.50 the expected F1-optimal decision boundary for a balanced dataset. The near-symmetric F1 profile around 0.50 ( at τ = 0.45 vs. 0.838 at τ = 0.55) confirms that 0.50 is the model’s intrinsic classification boundary, rather than an externally imposed value. This convergence between the F1-optimal threshold and the class-prior probability (0.50) is a structural property of balanced binary classification with BCE training, not a coincidence-and it provides an additional, independent justification for τ = 0.50: edges exceeding this threshold are simultaneously (a) distinguishable from random chance by construction, and (b) most informative for pathway reconstruction by calibration. Hyperparameter calibration for confidence threshold τ. Empirical validation of algorithmic robustness The following observations characterize the greedy pathway construction’s behavior on our specific graph structure. We distinguish three separate questions that must be answered independently before accepting greedy pathway construction as methodologically sound. Validation 1: does greedy extension degrade pathway quality? Motivation and scope. The greedy pathway construction (Algorithm 1) sequentially extends pathways by selecting the highest-probability next node. We examine whether this sequential strategy degrades pathway quality across extensions-a question we phrase as sequential quality stability, which we validate empirically on our specific graph structure. Unlike a formal guarantee, empirical validation is bounded by our calibrated probability distribution and does not transfer to other domains without re-examination. Motivation and scope. The greedy pathway construction (Algorithm 1) sequentially extends pathways by selecting the highest-probability next node. We examine whether this sequential strategy degrades pathway quality across extensions-a question we phrase as sequential quality stability, which we validate empirically on our specific graph structure. Unlike a formal guarantee, empirical validation is bounded by our calibrated probability distribution and does not transfer to other domains without re-examination. Observation 1 (monotonicity on top-3 pathways). Let be a pathway constructed by Algorithm 1 with edge probabilities , where τ = 0.50 is the F1-maximizing threshold selected on the validation set. Let S(P) be the geometric mean pathway score (Eq. 2). We examine whether the sequential extension preserves the inequality . On our constructed top-3 pathways (mean S(P) = 0.72), we observe this monotonicity pattern in 98.2% of extensions (1,891 out of 1,927 extensions across all top-3 pathways in the test set). The remaining 1.8% violations occur exclusively at extension steps where is close to the threshold (0.50 ≤ < 0.55), suggesting that the property is sensitive to the probabilistic noise near τ. Mathematical derivation: the monotonicity property holds when the new edge’s probability satisfies . To see this, let , so that . The inequality is equivalent to . Taking natural logarithms and rearranging yields . In our setting, all edges satisfy , so (geometric mean of numbers ). When S(P) is close to τ, the condition is only marginally stronger than , which explains why monotonicity holds in 98.2% of extensions despite the stricter theoretical condition. We therefore rely on empirical validation as the primary evidence, with the derivation above providing theoretical context. Calibration check: Table 10 reports that τ = 0.50 achieves F1 = 0.847 on the validation set. Because τ was selected by F1-maximization on the same validation set, this performance constitutes calibration verification rather than independent validation. To avoid circular reasoning, we separately report that on a held-out calibration set (20% of training data, not used for threshold selection), the mean S(P) of top-3 pathways is 0.74 ± 0.11, confirming that the threshold produces well-separated pathway scores even under distributional shift. Observation 1 (monotonicity on top-3 pathways). Let be a pathway constructed by Algorithm 1 with edge probabilities , where τ = 0.50 is the F1-maximizing threshold selected on the validation set. Let S(P) be the geometric mean pathway score (Eq. 2). We examine whether the sequential extension preserves the inequality . On our constructed top-3 pathways (mean S(P) = 0.72), we observe this monotonicity pattern in 98.2% of extensions (1,891 out of 1,927 extensions across all top-3 pathways in the test set). The remaining 1.8% violations occur exclusively at extension steps where is close to the threshold (0.50 ≤ < 0.55), suggesting that the property is sensitive to the probabilistic noise near τ. Mathematical derivation: the monotonicity property holds when the new edge’s probability satisfies . To see this, let , so that . The inequality is equivalent to . Taking natural logarithms and rearranging yields . In our setting, all edges satisfy , so (geometric mean of numbers ). When S(P) is close to τ, the condition is only marginally stronger than , which explains why monotonicity holds in 98.2% of extensions despite the stricter theoretical condition. We therefore rely on empirical validation as the primary evidence, with the derivation above providing theoretical context. Calibration check: Table 10 reports that τ = 0.50 achieves F1 = 0.847 on the validation set. Because τ was selected by F1-maximization on the same validation set, this performance constitutes calibration verification rather than independent validation. To avoid circular reasoning, we separately report that on a held-out calibration set (20% of training data, not used for threshold selection), the mean S(P) of top-3 pathways is 0.74 ± 0.11, confirming that the threshold produces well-separated pathway scores even under distributional shift. Validation 2: does greedy approximate beam search effectively? Motivation and scope. We compare greedy construction (k=1) against beam search with beam width k, which retains the top-k candidates at each step. The question is whether the greedy approximation produces comparable top-3 pathways to beam-3 on our data. This is an approximation quality question scoped to our graph structure, not a general convergence claim. Motivation and scope. We compare greedy construction (k=1) against beam search with beam width k, which retains the top-k candidates at each step. The question is whether the greedy approximation produces comparable top-3 pathways to beam-3 on our data. This is an approximation quality question scoped to our graph structure, not a general convergence claim. Observation 2 (greedy-beam equivalence for top-3). On the 2024 test graph, the mean probability separation between top-1 and top-2 candidates is , and the separation distribution is right-skewed (median Δ = 0.16, interquartile range [0.10, 0.24]; 14.3% of technology pairs have Δ < 0.08). This separation structure, combined with the multiplicative structure of S(P), produces the greedy-beam equivalence observed in Table 16: for top-3 pathways, greedy and beam-3 pathways are identical in 100% of test years. We note that this equivalence holds by construction for k=1 (greedy itself) vs. k=3 (beam-3), and does not guarantee equivalence for other beam widths. The observation that large probability separation makes greedy near-optimal is consistent with theoretical results in structured prediction where the optimal action has sufficiently higher probability than alternatives49,50. Unlike deterministic optimal-substructure problems51, probability-based greedy selection in graph neural networks lacks formal optimality guarantees; our empirical validation establishes that the sufficient-separation condition holds for our specific graph. Observation 2 (greedy-beam equivalence for top-3). On the 2024 test graph, the mean probability separation between top-1 and top-2 candidates is , and the separation distribution is right-skewed (median Δ = 0.16, interquartile range [0.10, 0.24]; 14.3% of technology pairs have Δ < 0.08). This separation structure, combined with the multiplicative structure of S(P), produces the greedy-beam equivalence observed in Table 16: for top-3 pathways, greedy and beam-3 pathways are identical in 100% of test years. We note that this equivalence holds by construction for k=1 (greedy itself) vs. k=3 (beam-3), and does not guarantee equivalence for other beam widths. The observation that large probability separation makes greedy near-optimal is consistent with theoretical results in structured prediction where the optimal action has sufficiently higher probability than alternatives49,50. Unlike deterministic optimal-substructure problems51, probability-based greedy selection in graph neural networks lacks formal optimality guarantees; our empirical validation establishes that the sufficient-separation condition holds for our specific graph. Limitations of greedy-beam equivalence. Table 16 shows that for pathways ranked 4-6, greedy and beam-3 still share the same set of candidates but with reordered rankings. For pathways ranked 7-10, three out of four candidates differ between greedy and beam-3. This indicates that greedy-beam equivalence is rank-sensitive: it reliably holds for top-3 identification but degrades for lower-ranked candidates. We therefore recommend beam-3 for pathway ranking tasks requiring more than three candidates. Limitations of greedy-beam equivalence. Table 16 shows that for pathways ranked 4-6, greedy and beam-3 still share the same set of candidates but with reordered rankings. For pathways ranked 7-10, three out of four candidates differ between greedy and beam-3. This indicates that greedy-beam equivalence is rank-sensitive: it reliably holds for top-3 identification but degrades for lower-ranked candidates. We therefore recommend beam-3 for pathway ranking tasks requiring more than three candidates. Limitations of empirical validation We emphasize that Validations 1 and 2 address algorithmic behavior under our specific calibrated probability distribution, which reflects the topology and link structure of the Chinese fishery S&T co-occurrence graph. Three scope conditions apply: Domain specificity. The probability separation structure (Δ = 0.18) is a property of our graph; other technology domains may exhibit different separation distributions, which would affect the robustness of greedy-beam equivalence. Domain specificity. The probability separation structure (Δ = 0.18) is a property of our graph; other technology domains may exhibit different separation distributions, which would affect the robustness of greedy-beam equivalence. Rank sensitivity. Both validations confirm that greedy is equivalent to beam-3 for top-3 identification. For applications requiring more comprehensive pathway rankings, beam search with k > 3 is recommended. Rank sensitivity. Both validations confirm that greedy is equivalent to beam-3 for top-3 identification. For applications requiring more comprehensive pathway rankings, beam search with k > 3 is recommended. Survivorship conditioning. Validation 1 conditions on successfully constructed pathways; pathways that terminate early due to are excluded. Whether the monotonicity property would hold for these terminated pathways cannot be assessed from our data, as no extension step exists. Survivorship conditioning. Validation 1 conditions on successfully constructed pathways; pathways that terminate early due to are excluded. Whether the monotonicity property would hold for these terminated pathways cannot be assessed from our data, as no extension step exists. Summary: taken together, the three-layer empirical validation (Monotonic Pathway Coherence, Greedy-Beam Equivalence, Expert Validation) provides convergent evidence that greedy pathway construction does not degrade pathway quality for top-3 identification in our specific graph structure. However, this evidence does not constitute formal mathematical guarantees; its transferability to other technology domains requires re-examination of the domain-specific scope conditions enumerated above. Experiments and analysis Experimental setup Environment setup The DTH-GNN model is implemented using PyTorch52, with heterogeneous graph operations handled by the Deep Graph Library (DGL) and temporal modeling by PyTorch Geometric, on a server equipped with an Intel Xeon Gold 6248R CPU and an NVIDIA RTX A6000 GPU. Key hyperparameters are as follows: learning rate = 0.001 with the Adam optimizer; epochs = 200 with early stopping (patience = 20) to prevent overfitting; batch size = 64; three graph convolutional layers; hidden dimension = 256. The loss function is binary cross-entropy. Computational cost: training DTH-GNN on the full temporal graph (15 snapshots, ∼1.1M edges in 2024) requires approximately 4.2 h on a single NVIDIA RTX A6000 GPU. Inference on the 2024 snapshot (computing all candidate link probabilities) requires approximately 12 min. For comparison, GCN training requires 1.8 h, R-GCN requires 2.5 h, and EvolveGCN requires 3.1 hours under identical hardware conditions. Hyperparameter search space: key hyperparameters were selected via grid search on the validation set (2023 snapshot) within the following ranges: learning rate ; hidden dimension ; number of R-GCN layers ; GRU hidden dimension ; dropout rate . The selected configuration (lr = 1e-3, hidden_dim = 256, layers = 3, GRU_dim = 256, dropout = 0.0) achieved the highest validation AUC. Early stopping with patience = 20 epochs was applied to prevent overfitting. Data splitting strategy To prevent future information leakage, the full temporal range (2010–2024) is divided into three non-overlapping periods: Training set: edges observed in 2010–2022 (inclusive, 13 time-step snapshots) Validation set: edges observed in 2023 (1 time-step snapshot, used for hyperparameter tuning and early stopping) Test set: edges observed in 2024 (1 time-step snapshot, used for final model evaluation) This split ensures that the model never observes future edges during training, which is consistent with the real-world decision-making scenario where planners must predict future connections based on historical evidence only. For each positive technology co-occurrence edge, five candidate negative edges are randomly selected from node pairs that do not co-occur in the same time-step snapshot. The 5-candidate filter reduces the probability of accidentally sampling weak-positive pairs as negative samples. From these five candidates, one negative edge is randomly retained to maintain a 1:1 positive-to-negative ratio across all datasets (training, validation, and test), ensuring temporal comparability while preventing class imbalance during both training and evaluation. Sample statistics Following PRISMA protocol screening of raw multi-source data retrieval, the final source corpus comprises 32,847 academic publications, 8,956 invention patents, and 1,856 research projects (totaling 43,659 records). These source records and the graph nodes reported in Table 7 use different accounting units. Literature nodes (A) represent graph-instantiated publications and patents only, while research projects are represented separately as project nodes (P). In the 2024 snapshot, the graph contains 33,862 literature nodes and 1,431 project nodes, together representing 35,293 document/project nodes. The remaining nodes are extracted entities: 29,876 technology nodes, 2,156 institution nodes, and 15,678 researcher nodes. Therefore, the total 2024 graph-node count (83,003) exceeds the source-record count because extracted technology, institution, and researcher entities are modeled as additional nodes, while the literature-node count alone is not expected to equal the total number of source records. The 2024 snapshot contains 1,132,547 edges, with technology-technology co-occurrence edges accounting for approximately 65% of all edges. Temporally partitioned: the training set spans 13 annual snapshots (2010-2022); the validation set comprises one snapshot (2023); and the test set comprises one snapshot (2024). All datasets maintain a 1:1 positive-to-negative ratio for the link prediction task, with negative samples generated through the strategy of “randomly selecting one from five candidate non-co-occurring node pairs per positive edge”. The exact edge counts for each partition are determined by the technology co-occurrence frequencies within the respective time steps. Baseline models To evaluate the contribution of different network components, the following baseline models are compared. Differences among baselines are summarized in Table 11. Table 11Differences among baseline models.ModelGNN-basedHeterogeneityDynamic evolutionNode2VecNoNoNoGCNYesNoNoR-GCNYesYesNoEvolveGCNYesNoYesDTH-GNN (proposed)YesYesYes Differences among baseline models. DyFormer baseline (computational feasibility note): we additionally implemented DyFormer35-a state-of-the-art temporal graph Transformer-on a restricted subgraph (top 5,000 technology nodes by degree, 2019-2024 snapshots). DyFormer employs a full pairwise attention mechanism that scales as with the number of nodes35, which is computationally prohibitive for the full graph (, ∼1.1M edges). To achieve efficient training, DyFormer relies on subgraph-based node sampling strategies35, which may miss cross-domain interactions involving peripheral technology nodes. On the restricted subgraph, DyFormer achieves AUC = 0.901 ± 0.006 and AP = 0.894 ± 0.005 (5 seeds), compared to DTH-GNN’s AUC = 0.937 ± 0.003 on the same restricted subgraph. This confirms DTH-GNN’s advantage even when controlling for graph scale. GraphGPS as a baseline (not applicable): we do not include GraphGPS36 as a baseline because GraphGPS is designed for static graph representation learning53 and does not natively support temporal dynamics. GraphGPS was evaluated on 16 static graph benchmarks (molecular graphs, protein networks, code ASTs) and employs fixed adjacency matrices53, making it fundamentally inapplicable to our temporal link prediction task. This is an architectural design difference, not merely a computational limitation. The DTH-GNN vs. DyFormer comparison therefore represents the most relevant Transformer baseline evaluation for temporal pathway prediction. Evaluation metrics Quantitative evaluation of link prediction employs AUC (Area Under the ROC Curve) and AP (Average Precision). AUC measures overall ranking ability, with values closer to 1 indicating better performance; AP summarizes precision and recall jointly by ranking predicted positive samples. Qualitative evaluation assesses predicted pathways on two dimensions-rationality and novelty-with novelty focusing on whether predictions reveal cross-domain technology convergence opportunities undetectable by traditional methods, and rationality assessing technical feasibility and application value. Five senior fishery domain experts-including two researchers from the Chinese Academy of Fishery Sciences, two university professors, and one enterprise technical director-anonymously scored the Top-K predicted pathways using a modified Delphi method54 with a seven-point scale (1 = very low, 4 = moderate, 7 = very high) across three dimensions (technical feasibility, innovation novelty, application value). Ablation experiments were also conducted to evaluate each module’s contribution. Results and analysis Quantitative model performance comparison Quantitative evaluation results for the link prediction task are shown in Table 12. Table 12Performance comparison of different models on the technology link prediction task.ModelAUC (mean ± std)AP (mean ± std)Node2Vec0.682 ± 0.0040.712 ± 0.003GCN0.785 ± 0.0070.760 ± 0.006R-GCN0.836 ± 0.0050.815 ± 0.004EvolveGCN0.891 ± 0.0040.880 ± 0.004DTH-GNN (Proposed)0.934 ± 0.0020.928 ± 0.003 All values are mean ± standard deviation over 5 random seeds (42, 123, 2024, 7, 2025). DTH-GNN vs. EvolveGCN: ΔAUC = +0.043, p < 0.001 (paired t-test, Bonferroni-corrected).Significant values are in bold. Performance comparison of different models on the technology link prediction task. All values are mean ± standard deviation over 5 random seeds (42, 123, 2024, 7, 2025). DTH-GNN vs. EvolveGCN: ΔAUC = +0.043, p < 0.001 (paired t-test, Bonferroni-corrected). Significant values are in bold. As shown in Table 12, GNN-based models (GCN, R-GCN, EvolveGCN, DTH-GNN) significantly outperform the traditional static embedding method (Node2Vec), demonstrating GNNs’ superiority in capturing complex network structure information7. R-GCN, which accounts for heterogeneity, substantially surpasses homogeneous GCN, confirming that distinguishing different node and relation types is critical for understanding innovation networks8. Models incorporating dynamic evolutionary analysis (EvolveGCN, DTH-GNN) outperform their static counterparts (R-GCN), indicating that temporal dynamic characteristics of S&T innovation are key information for predicting future trends. DTH-GNN achieves the best performance on all metrics, as it simultaneously accounts for both heterogeneity and dynamism, most comprehensively learning the complex patterns of the fishery S&T innovation network. Performance comparison on restricted subgraph DyFormer35 requires attention computation, which is computationally prohibitive for the full graph (). We therefore implemented DyFormer on a restricted subgraph consisting of the top 5000 technology nodes by degree (2019–2024 snapshots). DTH-GNN was re-evaluated on the same restricted subgraph for a fair comparison (Table 13). Table 13Performance comparison between DyFormer and DTH-GNN on the restricted subgraph (top 5,000 technology nodes, 2019-2024).ModelAUCAPDyFormer (5 seeds)0.901 ± 0.0060.894 ± 0.005DTH-GNN (same subgraph)0.937 ± 0.0030.930 ± 0.003DTH-GNN (full graph, main result)0.934 ± 0.0020.928 ± 0.003ΔAUC (DTH-GNN vs. DyFormer)+0.036–p-value (paired t-test)< 0.001–All results are reported as mean ± standard deviation across 5 random seeds (42, 123, 2024, 7, 2025). The p-value is computed from a paired t-test comparing seed-by-seed AUC values Performance comparison between DyFormer and DTH-GNN on the restricted subgraph (top 5,000 technology nodes, 2019-2024). All results are reported as mean ± standard deviation across 5 random seeds (42, 123, 2024, 7, 2025). The p-value is computed from a paired t-test comparing seed-by-seed AUC values Computational constraints note DyFormer’s attention complexity makes full-graph evaluation infeasible: a single forward pass on requires approximately 66 GB of GPU memory, exceeding the capacity of standard research-grade GPUs (NVIDIA A100 40 GB can handle ). The restricted subgraph evaluation (|V| = 5000) is therefore a necessary compromise that nevertheless provides a valid architectural comparison, as both models are evaluated on identical node sets. Subgraph note The marginally higher AUC on the restricted subgraph (0.937 vs. 0.934) reflects denser connectivity among top-k nodes, a known pattern in scale-free networks7; the full graph remains the primary setting for comprehensive pathway discovery. Assessment of potential data leakage in performance comparison Given the high AUC of 0.934, we conducted a systematic assessment of potential data leakage sources. Temporal split integrity. All reported AUC values are computed under the temporal split (training: 2010-2022; test: 2024). The model never observes future edges during training, ruling out direct temporal leakage. Temporal split integrity. All reported AUC values are computed under the temporal split (training: 2010-2022; test: 2024). The model never observes future edges during training, ruling out direct temporal leakage. Feature initialization integrity. Node features (BERT embeddings, structural metrics) are computed from static metadata (publication text, institutional names, project titles) that are fixed at data collection time and do not encode future graph structure. The GRU hidden state for new nodes is initialized to zero; sensitivity analysis shows this contributes at most +0.002 AUC. Feature initialization integrity. Node features (BERT embeddings, structural metrics) are computed from static metadata (publication text, institutional names, project titles) that are fixed at data collection time and do not encode future graph structure. The GRU hidden state for new nodes is initialized to zero; sensitivity analysis shows this contributes at most +0.002 AUC. Multi-seed stability. DTH-GNN’s AUC varies by only ± 0.002 across 5 random seeds, confirming results are not artifacts of a favorable random initialization. Multi-seed stability. DTH-GNN’s AUC varies by only ± 0.002 across 5 random seeds, confirming results are not artifacts of a favorable random initialization. Comparative perspective. AUC = 0.934 ± 0.002 is comparable to reported link prediction AUCs in similar S&T network studies (AUC = 0.923 on patent co-classification networks55; AUC = 0.915 on citation networks56-both involving approximately comparable heterogeneous networks in terms of multi-relational structure). The multi-relational, multi-temporal structure of our heterogeneous graph provides richer predictive signal than simpler homogeneous networks. Comparative perspective. AUC = 0.934 ± 0.002 is comparable to reported link prediction AUCs in similar S&T network studies (AUC = 0.923 on patent co-classification networks55; AUC = 0.915 on citation networks56-both involving approximately comparable heterogeneous networks in terms of multi-relational structure). The multi-relational, multi-temporal structure of our heterogeneous graph provides richer predictive signal than simpler homogeneous networks. Validation of temporal modeling necessity via data splitting strategy comparison To rigorously validate that the observed performance gains of DTH-GNN primarily stem from its capacity to model temporal dynamics rather than merely capturing static structural patterns, we conducted a comparative evaluation using two distinct data splitting strategies: (i) the temporal split (chronological 2010-2022/2023/2024 for training/validation/test) employed in the main experiments, and (ii) a random split (edges randomly shuffled and partitioned 70%/15%/15% across all 15 annual snapshots, i.e., 2024 edges may appear in training/validation). This random split relaxes the temporal boundary to test whether model performance gains stem from temporal generalization or from exploiting an enlarged training set. Results are presented in Table 14. Critical findings: the results reveal three critical findings: First, DTH-GNN achieves the highest absolute performance under the temporal split (AUC = 0.934), the evaluation scenario that most closely reflects real-world deployment where future data is unavailable. Under temporal split, DTH-GNN outperforms EvolveGCN by ΔAUC = +4.3% (0.934 vs. 0.891), astatistically significant margin (paired t-test, p < 0.001). This primary comparison establishes DTH-GNN’s superior temporal modeling capacity. Second, the random-split comparison provides supporting evidence for this claim. All models improve under random splitting (where future edges are visible during training), but the magnitude of improvement differs markedly: DTH-GNN’s marginal gain is only +0.7%, whereas static baselines exhibit double-digit gains (GCN: +10.7%, R-GCN: +6.5%). This differential response confirms that DTH-GNN relies less on training-data diversity to achieve its performance, consistent with its explicit temporal encoding. Notably, EvolveGCN (+1.4%) falls between static models and DTH-GNN, confirming it captures some temporal dynamics but fewer than DTH-GNN. Third, the performance hierarchy between models changes markedly across split strategies. Under random split, the gap between DTH-GNN and static baselines narrows considerably (e.g., DTH-GNN vs. R-GCN: absolute AUC gap = 0.040 under random split vs. 0.098 under temporal split; relative gap = 4.4% vs. 11.7%).Conversely, under temporal split-reflecting real-world deployment scenarios where future data is unavailable-DTH-GNN’s advantage becomes pronounced and statistically significant. This pattern confirms that temporal dynamics constitute significant discriminative information for accurate hot-spot pathway prediction in fishery S&T innovation networks, as further supported by the ablation experiment showing that removing the temporal encoding module (TEM) causes the largest single-module performance decline (+0.049 AUC). Interpretation of the temporal-random performance gap: a critical observation from Table 14 is that all models achieve higher AUC under random split than under temporal split. This is expected and does not indicate method failure. We clarify the logic below: Random split provides an upper-bound estimate. When edges are randomly shuffled across temporal snapshots, the model may observe “future” edges during training. This temporal leakage artificially inflates AUC. The random-split AUC represents the method’s capability ceiling under idealized conditions, not its realistic performance. Random split provides an upper-bound estimate. When edges are randomly shuffled across temporal snapshots, the model may observe “future” edges during training. This temporal leakage artificially inflates AUC. The random-split AUC represents the method’s capability ceiling under idealized conditions, not its realistic performance. Temporal split provides a realistic estimate. Under temporal split, the model operates under genuine forecasting conditions: only historical edges (2010-2022) are available for training. The temporal-split AUC (0.934 for DTH-GNN) represents the method’s actionable performance in real-world deployment, where future data is unavailable. Temporal split provides a realistic estimate. Under temporal split, the model operates under genuine forecasting conditions: only historical edges (2010-2022) are available for training. The temporal-split AUC (0.934 for DTH-GNN) represents the method’s actionable performance in real-world deployment, where future data is unavailable. The gap magnitude reflects temporal modeling sophistication. DTH-GNN exhibits the smallest gap (+0.7%), whereas static baselines exhibit double-digit gaps (GCN: +10.7%, R-GCN: +6.5%). This indicates that DTH-GNN relies less on temporal leakage to achieve high performance, confirming its superior temporal modeling. The gap magnitude reflects temporal modeling sophistication. DTH-GNN exhibits the smallest gap (+0.7%), whereas static baselines exhibit double-digit gaps (GCN: +10.7%, R-GCN: +6.5%). This indicates that DTH-GNN relies less on temporal leakage to achieve high performance, confirming its superior temporal modeling. Lower AUC under temporal split is expected, not a weakness. The temporal split explicitly removes the information advantage of future-edge visibility. A lower AUC under stricter evaluation conditions is a methodological virtue, not a flaw. The appropriate comparison is across methods under the same split, not across splits for the same method. Lower AUC under temporal split is expected, not a weakness. The temporal split explicitly removes the information advantage of future-edge visibility. A lower AUC under stricter evaluation conditions is a methodological virtue, not a flaw. The appropriate comparison is across methods under the same split, not across splits for the same method. Summary: the temporal-random gap is a methodological design choice, not a logical contradiction. DTH-GNN’s small gap (+0.7%) demonstrates that its performance stems from genuine temporal modeling, not from exploiting temporal leakage. Causal interpretation of performance differentials: beyond the quantitative performance comparison presented above, we provide a qualitative interpretation of why temporal dynamics contribute significantly to hot-spot pathway prediction in fishery S&T innovation. Argument 1: domain-specific temporal clustering: fishery S&T innovation exhibits pronounced temporal clustering of related technologies. For example, gene editing applications in aquaculture clustered around 2019-2022 following the democratization of CRISPR tools; IoT and sensor network applications surged after 2016 driven by smart agriculture national policies; and blue carbon and carbon-sink fisheries research accelerated post-2020 aligned with China’s carbon neutrality commitments. Static models that pool observations across all time steps cannot capture these regime-specific patterns, leading to averaging artifacts that dilute predictive signal. DTH-GNN’s explicit temporal encoding enables the model to recognize that a gene editing-disease resistance link formed in 2021 carries different predictive weight than the same link formed in 2015. Argument 2: irreversibility of technology trajectories: some technology transitions in fishery S&T are lumpy and partially irreversible. The shift from extensive to intensive aquaculture, the adoption of factory-based recirculating systems, and the transition from capture-based to culture-based marine ranching represent path-dependent transitions where early adoption creates infrastructure lock-in, skill accumulation, and institutional reinforcement. Models that ignore temporal ordering cannot distinguish between reversible co-occurrence patterns (e.g., two technologies that rise and fall together due to funding cycles) and path-dependent technological shifts (e.g., the irreversible substitution of manual feeding by automated precision feeding systems). DTH-GNN’s directional temporal modeling captures this asymmetry: the probability of observing edge at t depends not merely on whether existed at any prior time, but on the specific temporal trajectory of ’s relationship evolution. Argument 3: information degradation in static aggregation. When edges from different time steps are pooled without temporal tags, the model receives a degraded signal where contemporaneous and historical relationships are confounded. Consider a technology pair that co-occurred frequently in 2010-2015 (establishment phase) but rarely in 2020-2024 (mature phase). A static model observes high aggregate co-occurrence and predicts high future linkage probability-yet this prediction is erroneous if the technology pair has entered a saturation or substitution phase. Conversely, a pair with low historical but rapidly accelerating recent co-occurrence signals emerging hot-spot formation that static models miss. DTH-GNN’s temporal encoding preserves the temporal ordering, enabling the model to distinguish “trending” connections (rising edge weights in recent time steps) from “established” ones (stable high weights) and “decaying” ones (falling weights). To quantify the contribution of temporal dynamics, we computed the mutual information between temporal features (node age, edge recency, temporal clustering coefficient) and link existence in the test set. Temporal features account for a substantial share of mutual information (31.8%, 95% CI: [28.4%, 35.1%]), comparable to structural features (29.1%, [26.2%, 32.0%]) and higher than attribute features (19.7%, [17.1%, 22.3%]), using the Kraskov-Stögbauer-Grassberger (KSG) estimator with k=5 nearest neighbors after feature standardization and PCA reduction of high-dimensional BERT embeddings. This evidence, combined with the substantial performance degradation of static models under temporal split, indicates that temporal dynamics are significantly correlated with link prediction accuracy in this domain. We note an important caveat: mutual information analysis establishes statistical association between feature categories and link existence, but does not establish causal primacy-that is, we cannot conclude from MI alone that temporal dynamics are the “primary cause” of accurate predictions. The 2.7-percentage-point gap between temporal (31.8%) and structural (29.1%) features is modest and falls within the overlapping 95% confidence intervals ([28.4%, 35.1%] vs. [26.2%, 32.0%]), indicating that the difference is not statistically distinguishable at conventional significance levels. The remaining 19.4% is reported as a non-additive residual due to overlap among feature groups and estimator limitations, not as a directly measured contribution of feature interactions or model nonlinearity. To more directly assess the causal contribution of temporal features, we rely on the ablation experiment (Table 15) and the data-splitting comparison (Table 14), which provide stronger evidence for the necessity of temporal modeling. Ablation experiments To verify the contribution of each component of DTH-GNN, systematic ablation experiments were conducted, progressively testing the temporal encoding module (TEM), heterogeneous graph aggregation module (HGA), and dynamic graph update module (DGU). Results are presented in Table 15 and Fig. 6. Visualization of DTH-GNN ablation experiment results. Ablation results show that each module contributes significantly. Adding the temporal encoding module (TEM) to GCN raises AUC from 0.785 to 0.845 (Δ = +0.060), confirming that temporal information is essential for hot-spot pathway prediction. Further adding the heterogeneous encoding mechanism (HGA) raises AUC to 0.895 (Δ = +0.050 over the TEM-equipped model), demonstrating that heterogeneous node type differentiation provides complementary discriminative information beyond temporal encoding alone.The dynamic graph update module (DGU) contributes a smaller incremental gain to 0.910 (+1.5%), yet remains statistically significant for long-term pathway prediction. The complete DTH-GNN, trained end-to-end, achieves AUC = 0.934. The sum of incremental module contributions (+0.060 + 0.050 + 0.015 = 0.125) accounts for 83.9% of the total performance gain (Δ AUC = 0.149). The remaining +0.024 (16.1%) arises from end-to-end joint optimization, which allows the modules to co-adapt during training rather than being fitted sequentially. While this confirms the presence of a modest optimization-level synergy-the end-to-end gain (+0.024) exceeds DGU’s individual contribution (+0.015) but is smaller than those of HGA (+0.050) and TEM (+0.060)-the dominant source of performance improvement remains the complementary architectural information captured by each module. This modular decomposition suggests that DTH-GNN’s accuracy gains are primarily attributable to the heterogeneous and temporal graph encoding design rather than to optimization artifacts. Hot-spot pathway identification results The top-3 highest-probability technology links predicted for 2025 and beyond were visualized and analyzed. By connecting these high-probability links, multiple potential hot-spot pathways were identified. Pathway 1: smart aquaculture and precision management: IoT → Sensor Networks → Intelligent Net Cages → Online Water Quality Monitoring → Precision Feeding Models → Growth Behavior Analysis → Disease Early-Warning Systems → Blockchain-based Aquatic Product Traceability This pathway exemplifies the deep empowerment of traditional aquaculture by information technology. Beginning with foundational data collection (IoT, sensors), it progresses to mid-level intelligent equipment (smart net cages) and real-time monitoring, advancing to algorithm-driven decision-making (precision feeding, disease early warning), and culminating in product value enhancement (blockchain traceability). The model successfully captures the technological deepening logic from “sensing” to “decision-making” to “trust.” Pathway 2: biotechnology-driven green seed industry innovation: Genomics → Molecular Marker-Assisted Breeding → Gene Editing (CRISPR) → Disease/Stress-Resistance Trait Improvement → High-Yield/Fast-Growing Variety Cultivation → Cell Factory/Cultured Aquatic Products → Synthetic Biology → Alternative Protein Sources This pathway reveals that biotechnology-especially frontier gene editing-is reshaping aquatic seed industries. Beginning with basic genome analysis and proceeding through advanced breeding to improve key economic traits, the pathway extends to disruptive “cellular agriculture” and alternative protein domains, signaling that future fisheries may serve not only as traditional aquatic product providers but as important sources of biomanufactured and novel food products-carrying strategic significance for long-term food security57. Pathway 3: ecological fisheries and resource restoration convergence: Marine Ranching → Artificial Reefs → Habitat Restoration Technology → Environmental DNA (eDNA) Monitoring → Fishery Resource Acoustic Assessment → Ecosystem Modeling → Carbon-Sink Fisheries → Blue Carbon Economy. This pathway reflects the convergence of production and ecology. Starting from engineered ecological restoration measures (marine ranching, artificial reefs) and combining advanced ecological monitoring techniques (eDNA, acoustic assessment), it advances to model-based systematic management (ecosystem modeling) and ultimately aligns with China’s carbon neutrality strategy, pointing toward the emerging “blue carbon economy” domain16. The model successfully identifies the value evolution pathway of fisheries expanding from single food production to ecosystem service functions. Counter-intuitive pathway discoveries and policy implications: Beyond the three dominant pathways, DTH-GNN identifies several counter-intuitive patterns that challenge conventional assumptions in fishery innovation studies: Fishery identity disruption (Pathway 2): The terminal nodes of Pathway 2 (“cell factory/cultured aquatic products,” “synthetic biology,” “alternative protein sources”) represent a paradigm shift rather than incremental improvement. Conventional fishery innovation literature assumes that aquaculture evolves along “capture → culture → intensive culture” trajectories58. Our finding suggests that future fishery R&D may pivot toward biomanufacturing platforms, implying that policy should explicitly include cellular aquaculture and synthetic biology as priority disciplines. Fishery identity disruption (Pathway 2): The terminal nodes of Pathway 2 (“cell factory/cultured aquatic products,” “synthetic biology,” “alternative protein sources”) represent a paradigm shift rather than incremental improvement. Conventional fishery innovation literature assumes that aquaculture evolves along “capture → culture → intensive culture” trajectories58. Our finding suggests that future fishery R&D may pivot toward biomanufacturing platforms, implying that policy should explicitly include cellular aquaculture and synthetic biology as priority disciplines. Ecological monetization (Pathway 3): The progression from “marine ranching” to “blue carbon economy” reveals that fishery innovation is diverging from pure production toward ecosystem service monetization. While ecological fisheries are widely discussed in the literature59, the carbon-market pathway is underexplored. This finding implies that policy should establish standardized monitoring, reporting, and verification (MRV) frameworks linking eDNA-based assessments to carbon credit certification, enabling fishery enterprises to monetize ecosystem services. Ecological monetization (Pathway 3): The progression from “marine ranching” to “blue carbon economy” reveals that fishery innovation is diverging from pure production toward ecosystem service monetization. While ecological fisheries are widely discussed in the literature59, the carbon-market pathway is underexplored. This finding implies that policy should establish standardized monitoring, reporting, and verification (MRV) frameworks linking eDNA-based assessments to carbon credit certification, enabling fishery enterprises to monetize ecosystem services. Non-linear technology adoption (Pathway 1): The transition from “disease early-warning systems” to “blockchain-based traceability” spans multiple technology generations, contradicting the assumption of stage-gate technology adoption in aquaculture. This suggests that targeted investments in digital infrastructure can enable non-linear jumps, potentially leapfrogging intermediate technology stages in developing regions. Non-linear technology adoption (Pathway 1): The transition from “disease early-warning systems” to “blockchain-based traceability” spans multiple technology generations, contradicting the assumption of stage-gate technology adoption in aquaculture. This suggests that targeted investments in digital infrastructure can enable non-linear jumps, potentially leapfrogging intermediate technology stages in developing regions. These counter-intuitive findings demonstrate that DTH-GNN captures not only dominant trends but also emerging paradigm shifts that may escape traditional expert-based foresight methods. Additional evidence: from edge predictions to pathway quality: We acknowledged that edge-level accuracy does not guarantee pathway-level coherence. Here we provide supplementary evidence that the greedy heuristic produces meaningful pathways: Semantic coherence by construction. Algorithm 1 operates exclusively on technology-technology edges within the technology sub-graph. The DTH-GNN model learns embeddings that are conditioned on the full heterogeneous graph structure (including institution, project, and researcher nodes), meaning that the predicted link probabilities between technology pairs encode not only direct co-occurrence patterns but also indirect associations mediated by shared institutional affiliations, collaborative networks, and funding relationships. This ensures that predicted edges reflect innovation-driven associations rather than spurious statistical correlations. Semantic coherence by construction. Algorithm 1 operates exclusively on technology-technology edges within the technology sub-graph. The DTH-GNN model learns embeddings that are conditioned on the full heterogeneous graph structure (including institution, project, and researcher nodes), meaning that the predicted link probabilities between technology pairs encode not only direct co-occurrence patterns but also indirect associations mediated by shared institutional affiliations, collaborative networks, and funding relationships. This ensures that predicted edges reflect innovation-driven associations rather than spurious statistical correlations. Sequential reinforcement property empirically validated. The greedy construction imposes a monotonicity constraint: each successive edge in the pathway must exceed the confidence threshold τ. Empirically, this means that the geometric mean score S(P) of constructed pathways is well above τ: the three identified hot-spot pathways have mean edge probabilities of 0.72 (Smart Aquaculture), 0.68 (Green Seed Industry), and 0.65 (Ecological Fisheries), indicating that the greedy algorithm selects consistently strong edges rather than relying on a single high-probability edge. Sequential reinforcement property empirically validated. The greedy construction imposes a monotonicity constraint: each successive edge in the pathway must exceed the confidence threshold τ. Empirically, this means that the geometric mean score S(P) of constructed pathways is well above τ: the three identified hot-spot pathways have mean edge probabilities of 0.72 (Smart Aquaculture), 0.68 (Green Seed Industry), and 0.65 (Ecological Fisheries), indicating that the greedy algorithm selects consistently strong edges rather than relying on a single high-probability edge. Expert evaluation as primary validation. The ultimate test of pathway quality is whether domain experts recognize the hot-spot pathways as meaningful technological trajectories. The expert evaluation (Krippendorff’s α = 0.804, “almost perfect agreement”) provides this validation, with all three pathways receiving mean overall scores above 5.5 on a 7-point scale. Critically, experts assessed not only individual edge rationality but also pathway coherence-whether the sequence of technologies represents a logical evolutionary trajectory. The high coherence scores (Technical Feasibility: 5.87 ± 0.64; Application Value: 6.13 ± 0.35) confirm that the greedy construction produces interpretable pathways. Expert evaluation as primary validation. The ultimate test of pathway quality is whether domain experts recognize the hot-spot pathways as meaningful technological trajectories. The expert evaluation (Krippendorff’s α = 0.804, “almost perfect agreement”) provides this validation, with all three pathways receiving mean overall scores above 5.5 on a 7-point scale. Critically, experts assessed not only individual edge rationality but also pathway coherence-whether the sequence of technologies represents a logical evolutionary trajectory. The high coherence scores (Technical Feasibility: 5.87 ± 0.64; Application Value: 6.13 ± 0.35) confirm that the greedy construction produces interpretable pathways. Pathway construction algorithm: ablation and comparison To assess the robustness of hot-spot pathway to the construction algorithm, we compared three variants: (i) greedy sequential search (Algorithm 1), (ii) beam search with k=3, and (iii) beam search with k=5. Key finding: The top-3 pathways are fully stable across all three construction algorithms, confirming that the primary policy recommendations are robust to algorithmic choice. The greedy algorithm is retained as the default due to its favorable trade-off between computational efficiency (12 min vs. 312 min for beam-5) and pathway quality. Justification of greedy choice: The greedy sequential strategy is retained for the following reasons: (1) top-3 stability is confirmed; (2) deterministic outputs enhance interpretability for policy use; (3) the primary use case (R&D prioritization) prioritizes speed and clarity over exhaustive search of exponentially many pathways; and (4) the beam-5 expansion to all seed nodes is computationally prohibitive on the full graph. Observational patterns in cross-domain convergence of pathways identification While not a formal hypothesis requiring validation, we note an empirical pattern emerging from the pathway analysis: the three identified hot-spot pathways exhibit distinct convergence dynamics across technology domains (Table 17). Table 17Cross-domain technology co-occurrence patterns across hot-spot pathways.PathwayPrimary domain intersectionConvergence patternPathway 1: Smart aquacultureICT ↔ Traditional aquacultureGradual integrationPathway 2: Green seed industryBiotechnology ↔ Agricultural scienceAccelerated convergencePathway 3: Ecological fisheriesMarine science ↔ Environmental engineeringSystematic integration Cross-domain technology co-occurrence patterns across hot-spot pathways. This observational analysis provides descriptive insights into technology convergence dynamics, complementing the formal hypothesis testing of H1 and H2. Expert evaluation of hot-spot pathways To validate the rationality and novelty of predicted pathways, the three top-scoring pathways were selected and evaluated by five senior fishery S&T domain experts via a modified Delphi method. All participating experts were informed of the study objectives, procedures, and the intended use of their evaluations prior to participation, and written informed consent was obtained from each expert before the evaluation commenced. Expert scoring results are presented in Tables 18 and 19. Table 18Final anonymous scoring results for rationality of three predicted pathways.PathwayDimensionE1E2E3E4E5Mean ± SDSmart aquacultureTechnical feasibility676766.4 ± 0.55Innovation novelty565655.4 ± 0.89Application value677666.4 ± 0.55Overall score5.676.676.006.335.676.07 ± 0.43Green seed industryTechnical feasibility565565.4 ± 0.55Innovation novelty674655.6 ± 1.14Application value666666.0 ± 0.71Overall score5.676.335.005.675.675.73 ± 0.28Ecological fisheriesTechnical feasibility665665.8 ± 0.45Innovation novelty556555.2 ± 0.45Application value666666.0 ± 0.00Overall score5.675.675.675.675.675.67 ± 0.00 Final anonymous scoring results for rationality of three predicted pathways. Table 19Mean expert evaluation scores for predicted pathway rationality (all three pathways).DimensionMean ± SDMedianKrippendorff’s αAgreement levelTechnical feasibility5.87 ± 0.646.00.812Almost perfectInnovation novelty5.40 ± 0.655.00.756SubstantialApplication value6.13 ± 0.356.00.845Almost perfectOverall score5.82 ± 0.335.670.804Almost perfect Mean expert evaluation scores for predicted pathway rationality (all three pathways). Evaluation results show that the Krippendorff’s alpha coefficient60,61 was 0.804 for overall scores, reaching the “almost perfect agreement” level (), indicating good reliability of evaluation results. Alpha was computed using the interval-level distance metric on a matrix of n = 5 expert raters × 3 pathways, following the nominal-group consensus procedure described above.The smart aquaculture pathway scored highest (6.07 ± 0.43), with experts unanimously recognizing high technology maturity; the green seed industry pathway scored 5.73 ± 0.28, with initial divergence on innovation novelty converging after discussion; and the ecological fisheries pathway scored 5.67 ± 0.00 with the highest consistency, gaining recognition for policy alignment. Limitations include the small expert sample (n = 5) and regional limitation (all Mainland China experts). Future studies should expand sample size and incorporate international perspectives. Important methodological note: the reported AUC = 0.934 and AP = 0.928 are edge-level metrics for the link prediction task, not pathway-level metrics. Pathway identification quality depends on both individual edge prediction accuracy and the cumulative effect of sequential greedy construction. To assess pathway-level performance, we compute the pathway-level precision as the fraction of expert-validated “coherent” pathways among the top-K candidates: for K = 10, pathway-level precision is 0.70 (7 out of 10 pathways rated ≥5.0 overall; Krippendorff’s α = 0.687 for inter-rater agreement at K=10); for K = 5, pathway-level precision is 0.80 (4 out of 5 pathways rated ≥5.0 overall; α = 0.756); for K = 3, pathway-level precision is 1.00 (3 out of 3 pathways rated ≥5.0 overall; α = 0.804). These results indicate that while edge-level metrics substantially overestimate pathway-level accuracy (AUC = 0.934 vs. pathway precision = 0.70 at K=10), the top-ranked pathways (K ≤ 3) maintain high quality and near-perfect inter-rater agreement. This precision gradient has an important practical implication: DTH-GNN is best suited for prioritizing a small number of high-confidence pathways (top-3) rather than generating comprehensive pathway rankings at larger K values, where false positives become increasingly likely. Back-testing and external validation To assess whether DTH-GNN’s link prediction accuracy generalizes to unseen future time steps (rather than being an artifact of the 2024 test distribution), we conducted a retrospective back-testing experiment. Design: we trained DTH-GNN using data from 2010–2019 (inclusive, 10 annual snapshots) and used the trained model to predict technology co-occurrence links for the 2020 time step. Predictions were compared against actually observed links in 2020 publications, patents, and projects. Caveat: this design confounds two factors-reduced training data volume (10 vs. 13 snapshots in the primary experiment) and a longer prediction horizon (4 years vs. 1 year)-so the observed performance drop (ΔAUC = 0.087) should be interpreted as the joint effect of reduced training capacity and longer-term prediction uncertainty, and should not be attributed to temporal generalization limitations alone. Seed specification: the back-test uses 10 random seeds (10, 20, 30, 42, 100, 123, 456, 2024, 7, 2025) to increase robustness of the retrospective evaluation. The main experiment uses 5 seeds (42, 123, 2024, 7, 2025) for computational efficiency, consistent with standard practice in the GNN literature. Edge-level results: for the top-10% highest-probability predicted edges in 2020, DTH-GNN achieved mean AUC = 0.847 ± 0.003 and mean AP = 0.840 ± 0.004 across 10 independent training runs (Table 20), as well as precision = 0.734 and recall = 0.689 (Table 21). These results confirm robust temporal generalization that is not an artifact of a favorable random initialization. Table 20Edge-level link prediction performance of DTH-GNN on the 2020 back-testing hold-out set across 10 random seeds.SeedAUCAP100.8510.844200.8430.835300.8490.841420.8460.8381000.8520.8451230.8480.8404560.8440.83620240.8500.84370.8430.83420250.8470.840Mean ± SD0.847 ± 0.0030.840 ± 0.004 Edge-level link prediction performance of DTH-GNN on the 2020 back-testing hold-out set across 10 random seeds. Table 21Precision and recall of DTH-GNN on the 2020 hold-out set (deterministic, independent of random seed).MetricValuePrecision@Top-10% predicted edges0.734Recall@Top-10% predicted edges0.689Precision and recall are deterministic with respect to the trained model; they reflect the overlap between the model’s top-10% highest-probability predicted edges and the observed 2020 ground-truth edges, and do not vary across random seed initializations Precision and recall of DTH-GNN on the 2020 hold-out set (deterministic, independent of random seed). Precision and recall are deterministic with respect to the trained model; they reflect the overlap between the model’s top-10% highest-probability predicted edges and the observed 2020 ground-truth edges, and do not vary across random seed initializations Pathway-level back-test: we reconstructed the “Smart Aquaculture” pathway using only data available through 2019, and compared the predicted 2020 technology co-occurrence cluster against empirically observed co-occurrence clusters in 2020 data. The predicted pathway shared 7 out of 8 nodes with the empirically observed cluster (overlap ratio = 87.5%), providing direct evidence of pathway-level forecasting validity. Caveat: back-testing validates retrospective accuracy. Genuinely novel technology combinations (those without historical precedent) may not be detectable by this framework, as discussed in the Limitations section. Dynamic evolution of “smart fisheries”: a case study Focusing on the “smart fisheries” sub-domain, the dynamic evolution of internal technology themes is analyzed62 with attention to domain-specific factors unique to fishery science and technology. Biological cycle impacts. The “Smart Aquaculture” pathway is validated against the aquaculture production cycle. IoT sensor deployment peaks align with disease susceptibility periods (summer high-temperature seasons, approximately June-September). The model captures how technology adoption accelerates during high-risk periods when monitoring and automated intervention offer the highest marginal productivity gain for aquaculture operations. Resource constraint dynamics. The “Ecological Fisheries” pathway is contextualized against China’s fishery resource depletion timeline. The 2015-2020 marine fishing moratorium and the 2021 Yangtze River fishing ban created regulatory forcing that accelerated technology adoption in ecological restoration and resource monitoring. The model captures the technology response trajectory following these policy interventions, providing quantitative evidence of regulation-driven innovation patterns. Counter-intuitive finding. The model identifies “synthetic biology” as a high-probability node in Pathway 2 (Green Seed Industry) with a predicted edge probability of 0.71. This technology has not yet received mainstream policy attention in China’s fishery sector but emerges through indirect heterogeneous pathways (seed breeding → molecular biology institution → synthetic biology research collaboration). This represents a genuine counter-intuitive prediction that merits further policy attention. Smart fisheries technology grew rapidly from 2016 onward, driven by the convergence of AI and IoT. Biotechnology showed steady growth, while deep-sea fishing technology accelerated after 2020 under policy and market drivers. 2015: concept germination. The network’s core nodes are primarily basic information technologies-“IoT”, “RFID”, “sensors”-with dispersed application scenarios mainly focusing on point applications such as “pond aquaculture” and “water quality monitoring”. Network structure is relatively sparse. 2015: concept germination. The network’s core nodes are primarily basic information technologies-“IoT”, “RFID”, “sensors”-with dispersed application scenarios mainly focusing on point applications such as “pond aquaculture” and “water quality monitoring”. Network structure is relatively sparse. 2020: rapid development. The network becomes denser, with “big data”, “cloud computing”, “UAV”, and “remote sensing” emerging as new core nodes closely connected with existing technologies. Application scenarios extend to more complex systems such as “offshore deep-water aquaculture”, “precision fishing”, and “smart fishing harbors”. The emergence of “model”, “algorithm”, and “prediction” nodes marks the transition from data collection to intelligent decision-making. 2020: rapid development. The network becomes denser, with “big data”, “cloud computing”, “UAV”, and “remote sensing” emerging as new core nodes closely connected with existing technologies. Application scenarios extend to more complex systems such as “offshore deep-water aquaculture”, “precision fishing”, and “smart fishing harbors”. The emergence of “model”, “algorithm”, and “prediction” nodes marks the transition from data collection to intelligent decision-making. 2025 (forecast): deep integration and intelligence. “Artificial intelligence”, “digital twin”, “edge computing”, and “fishery large model” are predicted to become new and extremely important core nodes63. These technologies will integrate deeply with existing technology systems, forming a space-air-ground-sea integrated fishery perception and management system. Path predictions indicate that “digital twin fish farm” will connect “growth models”, “energy management”, and “market forecasting”, forming a highly intelligent production-operation loop. 2025 (forecast): deep integration and intelligence. “Artificial intelligence”, “digital twin”, “edge computing”, and “fishery large model” are predicted to become new and extremely important core nodes63. These technologies will integrate deeply with existing technology systems, forming a space-air-ground-sea integrated fishery perception and management system. Path predictions indicate that “digital twin fish farm” will connect “growth models”, “energy management”, and “market forecasting”, forming a highly intelligent production-operation loop. This dynamic evolution analysis clearly demonstrates the technological evolution paradigm of “smart fisheries” from “connectivity” to “platform” to “intelligence”, validating the proposed framework’s capacity to capture domain dynamic evolution patterns. Discussion Contributions and implications The results of this study demonstrate that the proposed framework, which constructs multi-source heterogeneous data into a dynamic heterogeneous graph and employs DTH-GNN for learning and prediction, is both feasible and effective. DTH-GNN comprehensively outperforms all baseline models in link prediction, confirming its superiority in capturing and modeling complex innovation networks8,9. These findings align with and extend prior work on dynamic network analysis for emerging technology identification13,20, with the key advance being the simultaneous modeling of heterogeneity and dynamism in a domain-specific context. The three identified high-potential pathways-smart aquaculture, green seed industry, and ecological fisheries-not only reveal internal technological deepening logic but, importantly, identify cross-domain technology convergence trends1,2. The smart aquaculture pathway, in particular, represents a comprehensive digital transformation trajectory that could substantially improve production efficiency and product traceability. The biotechnology pathway, culminating in synthetic biology and alternative proteins, signals a paradigm shift in how fisheries contribute to global food systems. The ecological fisheries pathway, connecting restoration technology with carbon-sink valuation, positions fisheries as active contributors to climate mitigation strategies. The case analysis of “smart fisheries” further demonstrates that the framework clearly and intuitively reveals the dynamic evolutionary process of internal structure within specific technology domains-from shifts in technology hotspots to transfers of core technology paradigms-providing a powerful analytical tool for understanding technology life cycles and anticipating future breakthroughs. From a practical standpoint,the hot-spot pathways provide indicative rather than prescriptive guidance for S&T planning and investment decisions. Specifically, the pathways highlight technology convergence directions that are statistically supported by historical innovation patterns, but they do not constitute causal predictions of inevitable technological futures. The translation from pathway identification to policy action requires additional considerations-including regulatory feasibility, market readiness, infrastructure requirements, and social acceptability-that lie beyond the scope of the current model. We therefore recommend that the framework be used as a screening and prioritization tool within a broader strategic planning process that incorporates qualitative expert judgment and contextual analysis, rather than as a standalone decision-making instrument. Correlation and causation: the identified hot-spot pathways are probabilistic co-occurrence sequences that reflect statistical regularities in historical innovation patterns. They do not constitute causal claims about inevitable technological evolution. A high predicted link probability indicates that two technologies have historically co-occurred frequently within similar institutional and collaborative contexts, but it does not guarantee that future co-occurrence will follow the same pattern. Several factors beyond what the model captures - including policy shifts, sudden scientific breakthroughs, market disruptions, and resource constraints - can alter technological trajectories in ways that the model cannot anticipate. We therefore recommend interpreting the hot-spot pathways as statistically grounded hypotheses for further investigation, not as deterministic forecasts. Algorithmic Limitations Of Greedy Pathway Construction. The greedy pathway construction algorithm (Algorithm 1) is an empirically validated heuristic, not a theoretically guaranteed optimization procedure. While three lines of empirical evidence support its practical effectiveness on our specific graph-(i) monotonic pathway coherence in 98.2% of extensions (Validation 1), (ii) greedy-beam equivalence for top-3 pathways (Validation 2), and (iii) pathway stability under edge perturbation (perturbation analysis)-these are data-dependent observations that hold for the Chinese fishery S&T co-occurrence network with its specific probability separation structure (). They do not constitute universal theoretical guarantees that would transfer to graphs with different topological or probability characteristics. Specifically, for graphs where the probability separation between top-1 and top-2 candidates is small (Δ < 0.08, which accounts for 14.3% of technology pairs in our graph), the greedy algorithm may produce suboptimal pathways. This rank-sensitivity is confirmed by our observation that greedy-beam equivalence degrades for pathways ranked 4–10 (Table 16). We therefore recommend that practitioners: (a) use DTH-GNN for prioritization of top-3 technology pathways rather than comprehensive ranking of all candidates; and (b) validate predicted pathways with domain experts before incorporating them into policy decisions, as our expert evaluation (Krippendorff’s α = 0.804) provides independent confirmation of pathway meaningfulness. Theory-method correspondence boundary: returning to the epistemological framework established in the theoretical section: our Dosi-inspired design heuristics provide structural guidance (what graph structure to build, what temporal encoding to use) but do not derive or prove the existence of the identified pathways. In other words, the Dosi framework motivates the architecture of our approach (heterogeneous nodes reflecting actor diversity, temporal modeling reflecting path-dependence), while the existence and content of specific pathways are empirical findings that stand on their own methodological merits (high link prediction accuracy, expert validation). This distinction is important for interpreting our theoretical contribution: we do not claim that Dosi’s theory “predicts” the Smart Aquaculture pathway, but rather that Dosi’s framework provides a principled basis for designing the computational infrastructure that discovers such pathways. Limitations and boundary conditions First, our empirical analysis focuses exclusively on China’s fishery S&T ecosystem, and the hot-spot pathways reflect domain-specific structural features (e.g., the dominant role of government-funded projects, the regulatory-driven adoption patterns characteristic of China’s central planning system) that may not transfer to market-driven innovation systems. The model’s architectural design (DTH-GNN) is domain-agnostic and transferable, but the learned parameters and hot-spot pathway are specific to the Chinese fishery context. Generalizability to other national contexts, other agricultural sub-sectors, or other technology domains requires systematic cross-contextual validation that we leave to future work. Accordingly, all policy recommendations should be understood as applicable to the Chinese fishery S&T planning context specifically, rather than as universal claims about global fisheries. Second, the model relies on Chinese-language databases (CNKI, Wanfang, CNIPA), which may underrepresent international collaboration patterns. Third, the five expert evaluators are all based in mainland China; incorporating diverse international perspectives could enhance the external validity of pathway validation. Future research should extend this framework to comparative cross-national studies and integrate multilingual data sources. Fourth, A fundamental limitation of the proposed pathway identification framework is its reliance on high-centrality seed nodes. The degree threshold () ensures that hot-spot pathway originate from technologies with established co-occurrence patterns, but this design systematically excludes emerging technologies with currently low degree but potentially high future impact. This bias is a deliberate trade-off: it prioritizes pathway reliability (based on sufficient historical co-occurrence evidence) over novelty (detecting emerging-from-scratch trajectories). The most transformative innovations-those arising from radical recombination of previously uncorrelated technologies-may not be detectable through this framework. Future work could relax the seed selection criterion (e.g., using citation burst detection or PageRank growth rate) to include emerging technologies with high acceleration but low absolute degree. Domain-specific validation and policy implications Expert validation of pathway predictions: to ensure the domain validity of identified pathways, we conducted a multi-stage expert evaluation following a modified Delphi method (Tables 18 and 19). Table 19 summarizes the validation results across three criteria. Interpretation: five fishery domain experts (including researchers from the Chinese Academy of Fishery Sciences, university professors, and enterprise technical directors) independently evaluated the Top-3 predicted pathways. The substantial inter-rater agreement (Krippendorff’s α = 0.804) and the pathway-level back-testing validation demonstrate that 87.5% of predicted pathway nodes overlap with empirically observed technology convergence in the 2020 holdout period. These results confirm that the identified pathways reflect meaningful domain patterns rather than artifacts of model optimization. Domain-specific characteristics. The identified pathways reflect domain-specific structural features of China’s fishery S&T ecosystem: Government-funded dominance. The fishery S&T network is predominantly shaped by national funding programs (NSFC grants, ministry-led projects), reflecting China’s central planning system in S&T investment. This structural feature explains the strong alignment between predicted pathways and policy documents (Table 22). Government-funded dominance. The fishery S&T network is predominantly shaped by national funding programs (NSFC grants, ministry-led projects), reflecting China’s central planning system in S&T investment. This structural feature explains the strong alignment between predicted pathways and policy documents (Table 22). Regulatory-driven adoption patterns. Technology adoption in fisheries follows regulatory timelines (e.g., fishing bans, environmental standards), which creates predictable temporal patterns that DTH-GNN successfully captures. Regulatory-driven adoption patterns. Technology adoption in fisheries follows regulatory timelines (e.g., fishing bans, environmental standards), which creates predictable temporal patterns that DTH-GNN successfully captures. Knowledge intensity and tacit learning. Fishery innovations require domain-specific tacit knowledge (e.g., aquaculture techniques, ecosystem monitoring), which manifests as strong institutional co-location in the co-occurrence network. This limits the transferability of identified pathways to contexts with different institutional structures. Knowledge intensity and tacit learning. Fishery innovations require domain-specific tacit knowledge (e.g., aquaculture techniques, ecosystem monitoring), which manifests as strong institutional co-location in the co-occurrence network. This limits the transferability of identified pathways to contexts with different institutional structures. Table 22Policy implications of identified hot-spot pathways.PathwayKey policy implicationTarget stakeholderSmart aquacultureInvest in digital infrastructure to enable non-linear technology adoption; include IoT and blockchain in R&D prioritiesMinistry of AgricultureGreen seed industryExplicitly include cellular aquaculture and synthetic biology as priority disciplinesS&T Planning bureausEcological fisheriesEstablish MRV frameworks linking eDNA assessments to carbon credit certificationEnvironmental agencies Policy implications of identified hot-spot pathways. Policy implications and recommendations: The hot-spot pathways provide indicative guidance-statistically supported technology convergence directions that require additional contextual considerations for policy translation. We recommend a three-stage decision framework: Screening (pathway identification). Use DTH-GNN to prioritize technology convergence directions based on historical co-occurrence strength and temporal dynamics. Screening (pathway identification). Use DTH-GNN to prioritize technology convergence directions based on historical co-occurrence strength and temporal dynamics. Assessment (expert validation). Validate prioritized pathways with domain experts, incorporating regulatory feasibility, market readiness, and infrastructure requirements. Assessment (expert validation). Validate prioritized pathways with domain experts, incorporating regulatory feasibility, market readiness, and infrastructure requirements. Integration (strategic planning). Incorporate validated pathways into broader S&T planning processes alongside qualitative expert judgment and contextual analysis. Integration (strategic planning). Incorporate validated pathways into broader S&T planning processes alongside qualitative expert judgment and contextual analysis. Limitations of policy applicability. Policy recommendations are specifically applicable to China’s fishery S&T planning context. Generalizability to other national contexts, other agricultural sub-sectors, or market-driven innovation systems requires systematic cross-contextual validation that we leave to future work. Conclusions This study proposes a comprehensive framework integrating GNN with dynamic evolutionary analysis for identifying hot-spot pathways in fishery S&T projects, making three principal contributions: (1) Methodological innovation: This study is, to our knowledge, among the first to systematically apply temporal heterogeneous GNNs to fishery S&T hot-spot pathway identification, introducing the DTH-GNN model capable of simultaneously processing multi-type node relations and temporal dynamic evolution. This method provides a new pathway prediction and analysis approach for existing hot-spot pathway identification research, filling the gap in GNN applications to fishery S&T innovation analysis. (2) Empirical contribution: The research findings-that the fishery S&T innovation network exhibits power-law degree distribution, prominent multi-actor co-evolution characteristics, and stage-wise leaping patterns in technology convergence-provide empirical evidence for applying innovation systems theory to specific industrial domains. (3) Practical value: The framework successfully identifies several high-potential hot-spot pathways covering smart aquaculture, green seed industry, and ecological fisheries, providing data-driven reference evidence for S&T investment prioritization and policy planning within the Chinese fishery context. Several limitations and future research directions merit attention: (1) Data breadth and depth-future work should incorporate additional data dimensions such as enterprise investment data, S&T policy texts, and market sales data. (2) Model interpretability-GNN “black-box” characteristics limit result interpretability; future work could combine graph attention visualization and counterfactual explanation methods to enhance transparency. (3) Causal inference-the hot-spot pathways are primarily correlation-based predictions rather than strict causal relationships; future work could introduce causal inference theory combined with quasi-experimental designs. (4) Human-machine interactive decision support systems-the ultimate goal is to develop an interactive human-machine collaborative decision support system integrating “data + algorithms + expert knowledge” for more efficient and scientific S&T innovation strategy management. In summary, leveraging GNNs and dynamic evolutionary analysis to empower fishery S&T hot-spot pathway identification represents a challenging but broadly promising research direction. This study takes a solid step forward and aspires to contribute intelligence and insights toward advancing sustainable fishery development in China and globally. Acknowledgements The authors sincerely thank the five fishery domain experts who participated in the Delphi evaluation for their invaluable time and insights. Author contributions Conceptualization, F.Z. and J.H.; methodology, F.Z.; software, F.Z.; validation, F.Z., J.H. and X.W.; formal analysis, F.Z.; investigation, F.Z.; resources, J.H.; data curation, F.Z.; writing-original draft preparation, F.Z.; writing-review and editing, J.H. and X.W.; visualization, F.Z.; supervision, J.H.; project administration, J.H.; funding acquisition, J.H. All authors have read and agreed to the published version of the manuscript. Funding This research was funded by the Special Fund for Basic Scientific Research Operations of the Chinese Academy of Fishery Sciences (Grant No. 2025A009). Data availability The processed graph data and model code that support the findings of this study are available from the corresponding author upon reasonable request. The raw data sources used in this study are publicly accessible via CNKI (https://oversea.cnki.net/index/), Wanfang Data (https://www.wanfangdata.com.cn), the NSFC Project Retrieval System (https://kd.nsfc.cn), and Derwent Innovation Index via Web of Science (https://www.webofscience.com/wos/), subject to their respective terms of service. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The processed graph data and model code that support the findings of this study are available from the corresponding author upon reasonable request. The raw data sources used in this study are publicly accessible via CNKI (https://oversea.cnki.net/index/), Wanfang Data (https://www.wanfangdata.com.cn), the NSFC Project Retrieval System (https://kd.nsfc.cn), and Derwent Innovation Index via Web of Science (https://www.webofscience.com/wos/), subject to their respective terms of service.