A multi-task learning-based fully connected neural network for personalized news recommendation Authors: ZhuoMin Ren, Hong Xie Published: 2026 ABSTRACT Traditional personalized news recommendation methods still face several limitations, such as inadequate modeling of dynamic user interests, difficulty in balancing accuracy and diversity, and significant performance degradation in cold-start scenarios. These limitations hinder their effectiveness in real-world applications. To address these issues, a Personalized News Recommendation Model via a Fully Connected Neural Network (MT-FCNN) is proposed. The model utilizes user behavior sequence embeddings to capture the dynamic evolution of user interests and adopts a multi-task learning framework that jointly optimizes click intention prediction and interest distribution learning, thereby improving recommendation performance in a unified representation space. The proposed model is evaluated on the publicly available Microsoft News Dataset and compared with established approaches, including Collaborative Filtering, Content-Based Recommendation, GRU4Rec, and Transformer-based models. Experimental results show that MT-FCNN improves AUC, NDCG@5, and CTR by 12.4%, 10.7%, and 8.9%, respectively. For cold-start users, recommendation accuracy (Precision@5) improves by 11.2%. Repeated experiments and statistical tests further confirm the model’s stability and significant advantages. This integrated framework introduces a new methodological perspective for recommendation systems while offering a more efficient technological approach for modeling user behavior in dynamic environments.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1038/s41598-026-57398-1. Introduction Personalized news recommendation systems play a pivotal role in information filtering by alleviating information overload. The rapid growth of digital content often overwhelms users, making it difficult individual preferences. Consequently, traditional recommendation systems are significantly in handling such complexity1,2. Developing efficient personalized news recommendation models has thus emerged as a central goal, with emphasis on accurately meeting individual needs by adapting to users’ evolving interests and behaviors3. Several challenges persist in mainstream approaches to personalized news recommendations. The absence of robust mechanisms for modeling the dynamic evolution of user interests, coupled with restricted recommendation accuracy, remains a significant obstacle. Furthermore, suboptimal performance in cold-start scenarios exacerbates these limitations, undermining both system efficacy and user satisfaction4. Overcoming these hurdles is essential to realizing the potential of recommendation systems to deliver precise, adaptive, and contextually relevant content. Traditional personalized recommendation methods, such as Collaborative Filtering (CF) and Content-Based Recommendation (CBR), provide foundational solutions but face notable limitations in complex recommendation scenarios. CF depends heavily on user-item interaction matrices and is therefore highly susceptible to data sparsity and cold-start issues. Its static nature also limits its capacity to adapt to evolving user preferences5. By contrast, CBR recommends items with similar content but relies on manually extracted attributes, which weakens its ability to capture the nuanced, multifaceted nature of user behavior and news content6. Deep learning approaches, including GRU4Rec models built on Recurrent Neural Networks (RNNs) and Transformer-based architectures, have brought substantial progress in sequential user behavior modeling. These models excel in analyzing temporal patterns in user interactions and provide advanced capabilities for predicting preferences. Nonetheless, single-objective optimization often constrains their performance, overlooking the advantages offered by multi-task learning frameworks that integrate multiple goals simultaneously7,8. Persistent challenges also remain in overcoming the cold-start problem and enhancing recommendation diversity, underscoring the need for further refinement in these modern methods9. To address challenges such as dynamic user behavior modeling, recommendation accuracy and diversity, and cold-start performance, this study proposes a Multi-Task Fully Connected Neural Network (MT-FCNN) for personalized news recommendation. The MT-FCNN framework incorporates embedding representations of user behavior sequences to capture the temporal evolution of user interests with greater precision. By leveraging a multi-task learning framework, the model jointly optimizes user click intention prediction and interest distribution modeling. This dual optimization strategy effectively balances the competing demands for recommendation accuracy and diversity while increasing resilience in cold-start conditions. In personalized news recommendation contexts, user interest distributions often exhibit high levels of complexity and fluidity. Accurately modeling such behaviors requires a system capable of adapting to these dynamic patterns. The MT-FCNN framework achieves this by integrating sophisticated behavioral modeling techniques with multi-task learning objectives, providing a robust and adaptable approach to recommendation generation. Through this combination, the framework addresses key limitations in existing methodologies and offers a significant advancement in the field of personalized news recommendation. In addition to improving recommendation performance, this study also examines the influence of personalized news recommendation models on user information retrieval behavior. As intelligent recommendation mechanisms increasingly become a primary channel for accessing news information, the accuracy, diversity, and dynamic adaptability of recommendation results directly affect users’ information acquisition efficiency, depth of content exploration, and reading behavior patterns. Therefore, investigating how recommendation models influence user information retrieval behavior is of considerable theoretical significance and can also provide practical guidance for optimizing intelligent information service systems. Focusing on the dual goals of accuracy and diversity in personalized news recommendations, this study emphasizes the importance of addressing users’ varied informational needs. Through experimental evaluation on the publicly available Microsoft News Dataset (MIND), the performance of the proposed MT-FCNN model is assessed, providing comprehensive empirical validation for advancements in this domain. The results reveal the MT-FCNN model’s superior performance over mainstream approaches, including CF, CBR, and deep learning models such as GRU4Rec and Transformer-based frameworks, across multiple critical evaluation metrics. The core contribution of this study is a novel recommendation framework, MT-FCNN, which integrates dynamic interest modeling with multi-task collaborative optimization. This innovation is reflected in three main aspects. First, the introduction of a dynamic sequence embedding paradigm overcomes the inherent limitations of traditional sequential models, such as RNNs and self-attention-based Transformers, in modeling long-term user behavior. By leveraging the nonlinear high-order interaction layers of a Fully Connected Neural Network (FCNN), the model directly captures semantic associations across time steps in user behavior (e.g., implicit topic shifts in news headlines), thereby avoiding the gradient vanishing issues of RNNs and the computational redundancy of Transformers. Second, the design of a multi-task collaborative architecture marks the first integration of Click-Through Rate (CTR) prediction and interest distribution learning under a unified optimization objective. This framework extracts common interest features through a shared behavioral embedding layer, while task-specific networks are used to separately optimize for immediate user intention and long-term preferences. This mechanism fundamentally addresses the “zero-sum dilemma” in traditional single-task models, where accuracy often comes at the expense of recommendation diversity, and vice versa. Third, the construction of a cold-start adaptive mechanism enables the effective utilization of sparse behavioral data. By incorporating multimodal feature fusion (including title semantics, news categories, and temporal cycles) and weakly supervised interest generalization, the model provides a reliable pathway for inferring the interests of new users with little or no historical interaction data. Together, these contributions offer both methodological advancement and practical value in enhancing the robustness, adaptability, and effectiveness of recommendation systems in dynamic and data-sparse environments. Compared to mainstream existing models, the MT-FCNN achieves multiple methodological breakthroughs. Traditional CF and CBR approaches are limited by static user profiling and manual feature engineering, making them ineffective in capturing the dynamic evolution of user interests. Moreover, they often fail in cold-start scenarios due to data sparsity. Although deep learning-based sequential models such as GRU4Rec and Transformer capture temporal dependencies, their single-task optimization frameworks tend to concentrate recommendations on popular content and thus sacrifice coverage of long-tail items. The innovation of MT-FCNN lies in several key aspects. In terms of dynamic modeling, it abandons the sequential recursion of RNNs and the self-attention mechanisms of Transformers. Instead, it employs densely connected FCNN layers to capture global semantic interactions across user behaviors. This approach eliminates the need for positional encoding and reduces the computational complexity associated with long sequences. In multi-task design, MT-FCNN is the first to combine the immediacy of click prediction with the breadth of interest distribution learning. By introducing adaptive task weighting (e.g., α = 0.5), the model dynamically balances task priorities based on user activity levels. For cold-start scenarios, a temporal decay multimodal feature fusion mechanism is introduced, enabling the model to generalize latent interest patterns from highly sparse behavioral data, rather than relying on the completion of explicit interaction matrices. These differentiated designs enable MT-FCNN to overcome the static modeling, single-objective focus, and strong data dependency inherent in traditional models. Consequently, it provides a new methodological framework that enhances the dynamic adaptability and diversity assurance of personalized recommendation systems. The structure of this study is organized as follows. Section 1 presents the introduction, outlining the research background and emphasizing the severity of the information overload problem as well as the significance of personalized news recommendation systems in the context of information filtering. This section also identifies the key challenges addressed in the study, clarifies the research objectives and significance, and summarizes the main contributions of the work. Section 2 provides a comprehensive review of related work, systematically examining the current state of research in personalized news recommendation. It covers traditional recommendation approaches, the application of deep learning techniques in recommendation systems, and recent advances in multi-task learning, thereby establishing the theoretical foundation and research context for this study. Section 3 elaborates on the design of the MT-FCNN model, including the overall architecture, the user behavior sequence embedding module, the multi-task learning framework, and the model’s training and optimization strategies. As the core of the study, this section aims to provide a comprehensive technical perspective for readers. Section 4 presents the experimental design, results analysis, and performance comparisons with existing mainstream models. Through empirical data, it demonstrates the effectiveness and advantages of the MT-FCNN model. Section 5 offers an in-depth discussion of the findings, exploring the theoretical mechanisms underlying the model’s performance, examining the practical implications, identifying limitations, and outlining potential directions for future research. Finally, Sect. 6 concludes the study by summarizing the main findings and highlighting the innovative value and application potential of the MT-FCNN model in the field of personalized news recommendation. Related work Traditional recommendation methods and feature modeling Personalized news recommendation systems have garnered significant attention in recent years, becoming a focal point in the field of information filtering. The growing interest from both academia and industry has spurred a variety of research efforts aimed at enhancing the efficacy of these systems. Traditional approaches, such as CF and CBR, have been foundational in early studies. For example, Wiratama et al. (2024) proposed a content-based approach that analyzed news content features to generate personalized recommendations10. However, this approach’s reliance on manually extracted features restricted its adaptability across different news domains. In parallel, Siyal et al. (2021) emphasized the significant influence of external factors, such as workplace dynamics and incentives, on individual motivation and behavior11. This insight offers valuable input for user behavior modeling within personalized recommendation systems. Specifically, the incorporation of external stimuli during information consumption—such as CTR and the format in which recommendations are presented—plays a crucial role in shaping users’ interaction patterns and interest distribution. Through the dynamic analysis of user behavior sequences and the application of multi-task learning, the influence of these external factors can be better captured, potentially leading to more accurate and contextually aware recommendation systems. Nevertheless, these traditional approaches primarily rely on static modeling and therefore struggle to capture the dynamic evolution of user interests. Ahmadian et al. (2025) provided a systematic overview of the application of Nonnegative Matrix Factorization (NMF) in recommendation systems. The authors emphasized that nonnegativity constraints on latent factors improved model interpretability, particularly when handling implicit feedback data such as click behaviors12. However, the inherent limitation of NMF lies in its linear factorization assumption, which fails to capture dynamic and nonlinear relationships within user behavior sequences, including interest drift and cross-category interactions. These challenges are specifically addressed by MT-FCNN, which leverages multi-layer fully connected networks and multi-task learning to model such complex dynamics more effectively. Deep learning and sequential modeling methods The advent of deep learning has revitalized recommendation systems, with numerous studies incorporating its capabilities into personalized recommendation models. Lin et al. (2024) introduced GRU4Rec, a deep recommendation model based on gated recurrent units (GRU), marking a significant milestone in the use of RNNs to dynamically model user behavior sequences. This approach improved both the timeliness and accuracy of recommendations13. Building on this, Rahman et al. (2024) advanced the field by employing the Transformer model, leveraging self-attention to capture global sequence patterns. This method has found widespread use in recommendation applications14. In addition, Lin et al. (2024) proposed a service recommendation method based on deep neural CF, which effectively integrated multi-source data and provided a solution for intelligent cloud-edge collaborative applications15. Its design for handling heterogeneous data sources offered valuable insights for recommendation systems operating in complex environments. Despite their success, these models often focus narrowly on optimizing a single objective, overlooking the need for diverse recommendation outcomes and the multidimensional nature of user interests. The growing interest in multi-task learning for recommendation systems has led to the development of more sophisticated models. Salamai et al. (2023) introduced a Multi-Task Learning framework for recommendation systems, aiming to optimize CTR and conversion rate predictions simultaneously, thereby enhancing generalization through parameter sharing16. This framework provides critical insights into multi-objective optimization in recommendation systems. Further contributions by Muzahid et al. (2020) improved recommendation performance through a multi-task learning framework using hard and soft parameter sharing; however, its capacity to model dynamic user interests remained limited17. Additionally, the role of embedding techniques in enhancing recommendation effectiveness has attracted increasing attention. For example, Patil et al. (2024) demonstrated the application of deep neural networks to embed user behavior features in Google’s video recommendation system, offering a practical model for large-scale recommendation environments18. These approaches primarily improve multi-objective optimization and representation learning capabilities. Through parameter sharing and embedding-based modeling, they enhance the fundamental representational capacity of recommendation systems. However, most of these methods still rely on static features or shallow interaction modeling, making it difficult to effectively capture the rapid temporal evolution of user interests in news recommendation scenarios. In addition, their limited ability to model higher-order nonlinear relationships within complex behavior sequences constrains their adaptability and generalization performance in real-world large-scale applications. Consequently, these methods mainly address the efficiency of representation learning rather than fully resolving the challenge of dynamic interest modeling. In the field of Graph Neural Networks (GNN) and structured representation learning, research has primarily focused on leveraging graph structures to model complex relationships between users and items, thereby improving recommendation quality and representational capability. El Alaoui et al. (2022) proposed an item recommendation system based on deep GraphSAGE, which addressed the over-smoothing problem in deep GNN by integrating jumping knowledge connections and ordinal aggregation networks19. El Alaoui et al. (2024) further introduced adaptive edge attributes into user–item interactions through a dynamic graph attention network. By jointly optimizing model parameters and edge features, the proposed framework significantly improved recommendation accuracy and novelty in contextual recommendation tasks20. In another study, El Alaoui et al. (2024) compared CF, CBR, and hybrid recommendation methods in scientific literature recommendation. Their findings showed that hybrid methods significantly outperformed other approaches in terms of accuracy, ranking quality, and novelty across two datasets, while also effectively alleviating cold-start and sparsity problems21. El Alaoui et al. (2025) combined GraphSAGE with capsule networks to generate integrated representations for each node from multiple perspectives within recommendation sessions, effectively capturing complex item transitions and cross-session behavioral patterns22. Graph-based recommendation methods exhibit clear advantages in modeling higher-order user–item relationships and can effectively capture complex interaction structures, leading to strong performance across various recommendation tasks. However, their effectiveness is highly dependent on the quality of graph construction. In dynamic news recommendation scenarios, graph structures require frequent updates, resulting in substantial computational overhead. In addition, these methods often struggle to support real-time online updates and generally exhibit limited scalability in high-frequency information stream environments. Consequently, although graph-based models enhance structural representation capability, a clear trade-off remains between dynamic adaptability and computational efficiency. On the other hand, the GNN framework proposed by Nejadshamsi et al. (2025) integrates geographic and semantic contextual information to model complex spatial associations23, offering new insights into constructing user-news interaction graphs within recommendation systems. Although GNN frameworks are effective in mining high-order collaborative signals—such as the propagation of group-level user interests—their reliance on explicitly constructed graph structures poses challenges for real-time recommendation scenarios. The temporal dynamics of user behavior sequences and the rapid update frequency of news content make static graph construction computationally intensive and less scalable. In contrast, MT-FCNN eliminates the need for predefined graph structures by directly embedding user behavior sequences and learning feature interactions through fully connected layers, making it more suitable for highly dynamic news recommendation environments. Furthermore, the spectral clustering optimization method introduced by Berahmand et al. (2025) offers theoretical support for user clustering in unsupervised settings24, but its reliance on labeled data stands in contrast to the end-to-end supervised learning approach adopted by MT-FCNN. These approaches provide important theoretical foundations for structured modeling, unsupervised learning, and complex relationship mining, and they improve the representational capability and robustness of recommendation systems to a certain extent. Nevertheless, most of these methods rely heavily on explicit graph structures or clustering assumptions, which introduce challenges such as high construction cost, difficulty in dynamic updating, and limited real-time adaptability in news recommendation scenarios. In addition, their end-to-end learning capability remains relatively limited, making it difficult to fully capture complex nonlinear interest evolution processes. By comparison, the proposed MT-FCNN directly models user behavior sequences through fully connected networks and integrates a multi-task learning mechanism to achieve dynamic interest modeling without requiring explicit graph structures. This design provides clear advantages in both scalability and real-time adaptability. Research on recommendation systems has achieved significant progress across multiple related domains, further expanding methodological boundaries and application scenarios. In the field of graph representation learning, Wang et al. (2025) introduced a self-supervised graph contrastive learning framework that enhanced model representation through advanced sampling strategies25. This development provided a new perspective for mining deep user associations from graph-structured data. Addressing resource constraints and privacy challenges in Internet of Things (IoT) environments, Wan et al. (2025) developed a lightweight and privacy-preserving IoT service recommendation scheme based on learning-to-hash techniques26. Its lightweight design and privacy considerations hold practical significance for recommendation applications on mobile and edge devices. In addition, Liu et al. (2024) investigated privacy-aware cross-platform social relationship prediction based on microservice architecture27. By leveraging sequential information to process cross-platform data, this approach offered inspiration for building scalable and privacy-preserving recommendation systems. Overall, these studies have advanced recommendation systems from the perspectives of self-supervised learning, lightweight privacy-preserving design, and microservice-driven architectures. The proposed MT-FCNN model adopts a distinct technical approach by directly modeling dynamic user behavior sequences through FCNN models and multi-task learning. Despite this difference, it aligns with these frontier studies in addressing core challenges such as data sparsity, improving model representational capacity, and balancing multi-objective optimization. Summary and comparison A comprehensive review of existing studies indicates that traditional recommendation methods, deep sequential models, multi-task learning approaches, and graph-based modeling techniques have each improved recommendation system performance from different perspectives. However, substantial limitations remain when attempting to address dynamic news recommendation within a unified framework. From the perspective of scalability, traditional CF and content-based methods rely heavily on static features and explicit similarity calculations, making them difficult to adapt to large-scale dynamic data streams. Although GNN-based approaches are capable of modeling higher-order relationships, they depend on complex graph construction and frequent graph updates, resulting in considerable computational overhead and limited real-time scalability in high-frequency news recommendation scenarios. Regarding temporal modeling capability, models such as GRU4Rec and Transformer-based architectures can capture certain sequential dependencies, but they are generally optimized for single-task objectives. As a result, they remain insufficient for modeling long-term user interest evolution, multi-objective behavioral changes, and cross-domain behavior transfer, making it difficult to fully reflect dynamic interest drift in real-world news consumption. In terms of interpretability, although NMF and some graph-based models provide a degree of structural interpretability, deep sequential models and multi-task neural networks largely remain “black-box” systems. These models lack a unified and stable mechanism for explaining prediction decisions, which limits their controllability and reliability in practical recommendation environments. From the perspective of data integration capability, existing approaches often process user behavior, content features, and structural relationships separately, rather than within a unified modeling framework. For example, multi-task learning methods mainly rely on shared representations to improve prediction performance, while graph-based methods focus on structural information fusion. Nevertheless, both approaches struggle to simultaneously integrate behavior sequences, semantic content information, and multi-objective task signals within a single framework. Based on the above analysis, existing methods still exhibit clear limitations in dynamic interest modeling, structural dependency, computational efficiency, and multi-source information integration. To address these issues, the proposed MT-FCNN directly models user behavior sequences through FCNN, thereby avoiding reliance on explicit graph structures. In addition, a multi-task learning mechanism is introduced to jointly optimize click prediction and interest distribution modeling. This design improves the model’s ability to capture dynamic user interests while maintaining computational efficiency and enhancing overall recommendation performance. Compared with existing approaches, the proposed framework demonstrates better adaptability and more balanced advantages in scalability, temporal modeling capability, and collaborative multi-task optimization. Research model MT-FCNN model construction To overcome the constraints of traditional personalized news recommendation systems—particularly capturing dynamic interest, ensuring diversity, and addressing cold-start scenarios— FCNN-based architecture is adopted. This choice is theoretically justified by its efficient modeling of high-order feature interactions, and practically justified by its ability to balance between dynamic interest modeling and resource consumption in multi-task recommendation settings. FCNN is specifically designed to address the complexity of high-order feature interactions embedded in user behavior sequences. By stacking multiple nonlinear fully connected layers, the model can flexibly capture the evolving patterns of user interests over time. More specifically, user behavior sequences exhibit both temporal dependencies and multi-level semantic associations—for instance, in news category, headline semantics, and time distribution. Traditional RNNs can capture temporal dependencies, but their recursive structure makes them vulnerable to vanishing gradients in long sequences and limits their ability to separate global context from local interest signals. Transformer-based models enhance global modeling through self-attention, but their computational complexity grows quadratically with sequence length. Transformer-based models, although enhancing global modeling through self-attention mechanisms, suffer from a quadratic growth in computational complexity with respect to sequence length. When the number of time steps exceeds 100, this complexity becomes a major bottleneck for training efficiency. In contrast, FCNN employs densely connected parameter matrices to perform high-order nonlinear transformations directly on the embedded representations of behavior sequences. This design overcomes RNNs’ limitations in modeling long-term dependencies and reduces the computational overhead introduced by Transformer architectures. Moreover, in a multi-task learning framework, the shared layers of FCNN are well-suited to extract common features for both CTR prediction and interest distribution learning, while the task-specific layers learn fine-grained representations for each objective. The study by Zuo et al. (2025) further confirmed that FCNN enhances the robustness of implicit interest modeling through effective feature crossing in news recommendation scenarios28, especially under cold-start conditions where user behavior is sparse. This demonstrates FCNN’s adaptability and efficiency, making it a compelling architecture for personalized recommendation systems operating under dynamic and resource-constrained environments. This study introduces the MT-FCNN model. Built on an FCNN architecture, this model incorporates a multi-task learning framework. The central concept revolves around embedding user behavior sequences and integrating them with a multi-task approach to jointly optimize both user click intention prediction and interest distribution modeling. This synergy results in a holistic improvement in recommendation system performance. The architecture of the MT-FCNN model is depicted in Fig. 1. Fig. 1MT-FCNN architecture. MT-FCNN architecture. The MT-FCNN architecture is structured around three primary modules, each contributing to the enhancement of recommendation accuracy and diversity via refined feature processing and multi-task optimization. The user behavior sequence embedding module plays a pivotal role in extracting features from the historical behavior sequence, converting them into high-dimensional semantic vectors. The shared FCNN further processes these embeddings, learning more complex representations through shared layers. The multi-task learning module encompasses two parallel tasks: user click intention prediction and user interest distribution learning, which are optimized together to improve model performance. At the input level, the user behavior sequence embedding module maps the historical click sequence of user u into high-dimensional semantic vectors. Denote the historical sequence of user u as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{n}_{1},{n}_{2},\cdots\:,{n}_{T}\right\}$$\end{document}, where T signifies the sequence length, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document} is the t-th news item. The embedding vector corresponding to each news item \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document} is represented as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}\in\:{\mathbb{R}}^{d}$$\end{document}, with the embedding process captured in Eq. (1):1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{E}=\left[{e}_{1},{e}_{2},\cdots\:,{e}_{L}\right]$$\end{document} In Eq. (1), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{E}\in\:{\mathbb{R}}^{L\times\:d}$$\end{document} denotes the embedding matrix of the user behavior sequence, where L represents the length of the historical behavior sequence (e.g., the number of news articles clicked by the user), and d is the dimension of each embedding vector (set to 128 in this experiment). The vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}\in\:{\mathbb{R}}^{d}$$\end{document} corresponds to the embedding of the t-th behavior (news click), constructed by concatenating the title embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}^{title}$$\end{document}, category embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}^{category}$$\end{document}, and time embedding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}^{time}$$\end{document}, i.e., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}=Concat({e}_{t}^{title},{e}_{t}^{category},{e}_{t}^{time})$$\end{document}. The title embedding is extracted using a pretrained Bidirectional Encoder Representations from Transformers (BERT) model, with dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{title}=64$$\end{document}. The category embedding is obtained by mapping a one-hot encoded category vector into a dense space of dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{category}=32$$\end{document}. The time embedding encodes periodic features such as hour of day and day of week through sinusoidal functions, with dimension \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{time}=32$$\end{document}. This multi-modal feature fusion design enhances the semantic completeness of behavior representations. To capture temporal dynamics and global feature interactions in user behavior, the embedding vectors are processed through multiple layers of the FCNN, as represented in Eq. (2):2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:H=f({W}_{1}E+{b}_{1})$$\end{document} In Eq. (2), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{1}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{1}$$\end{document} are trainable parameters, while f denotes the activation function. The shared FCNN module performs higher-order feature interactions and representation learning on the embedded features. Its primary objective is to enhance the expressive capacity of user behavior features through nonlinear transformations, simultaneously creating a shared feature space for the subsequent multi-task modules. Let the shared layer consist of fully connected layers, with the output of each layer described by Eq. (3):3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}^{\left(l\right)}=f({W}^{\left(l\right)}{h}^{\left(l-1\right)}+{b}^{\left(l\right)})$$\end{document} In Eq. (3), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}^{\left(l\right)}\in\:{\mathbb{R}}^{k}$$\end{document}represents the output features from the -th layer; \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}^{\left(l\right)}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}^{\left(l\right)}$$\end{document} represent the parameter matrix and bias vector; k denotes the dimension of the hidden layer. The initial input \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}^{\left(0\right)}=H$$\end{document} originates from the output feature of the embedding module. Through successive layers of the shared neural network, higher-order semantic representations of the user behavior sequence are learned, establishing a feature foundation for joint optimization in the following multi-task modules. The multi-task learning module incorporates two primary tasks: the user click intention prediction task and the user interest distribution learning task. The former predicts the likelihood of a user clicking on a news article, while the latter enhances recommendation diversity by capturing the distribution of user interests across various domains. The user click intention prediction task is formulated as a binary classification problem, optimized using the cross-entropy loss function, as depicted in Eq. (4):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{click}=-\frac{1}{N}\sum\:_{i=1}^{N}[{y}_{i}\mathrm{log}\left({\widehat{y}}_{i}\right)+\left(1-{y}_{i}\right)\mathrm{log}\left(1-{\widehat{y}}_{i}\right)]$$\end{document} In Eq. (4), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{y}_{i}$$\end{document} represents the true label, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{y}}_{i}$$\end{document} denotes the predicted probability. The user interest distribution learning task aims to minimize the prediction error in the user’s interest distribution, with the loss function defined as the mean squared error (MSE), shown in Eq. (5):5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{interest}=\frac{1}{N}\sum\:_{i=1}^{N}{\lVert{p}_{i}-{\widehat{p}}_{i}\lVert}_{2}^{2}$$\end{document} In Eq. (5), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{i}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{p}}_{i}$$\end{document} represent the true and predicted interest distributions, respectively. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\parallel\:\cdot\:{\parallel\:}_{2}$$\end{document} represents the Euclidean norm (L2 norm), and the superscript “2” denotes the squaring operation used in the computation of MSE. To combine these tasks, the total objective function for multi-task learning is the weighted sum of the individual task losses, expressed in Eq. (6):6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{total}=\alpha\:{\mathcal{L}}_{click}+(1-\alpha\:){\mathcal{L}}_{interest}$$\end{document} In Eq. (6), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:\in\:\left[\mathrm{0,1}\right]$$\end{document} is the task weight coefficient, controlling the relative importance of each task. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{\mathrm{c}\mathrm{l}\mathrm{i}\mathrm{c}\mathrm{k}}$$\end{document} represents the cross-entropy loss for the click prediction task, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{\mathrm{i}\mathrm{n}\mathrm{t}\mathrm{e}\mathrm{r}\mathrm{e}\mathrm{s}\mathrm{t}}$$\end{document} denotes the MSE loss for the interest distribution learning task. User behavior sequence embedding modeling The sequence of user behavior encapsulates dynamic patterns of interest and serves as a vital component in personalized news recommendation systems. In MT-FCNN, this sequence is transformed into high-dimensional vector representations through embedding. This process captures the evolution of user interests as well as the semantic connections between user actions and content. Let \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{u}=\left\{{n}_{1},{n}_{2},\cdots\:,{n}_{L}\right\}$$\end{document} denote the historical click sequence of user u, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document} is the news article clicked at time t, and L represents the sequence length. Each news article \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document} contains various features, including title, content, and category. To standardize the encoding of these features, embedding representations are employed, mapping them into a dense, high-dimensional space29. This process is captured in Eq. (7):7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}={f}_{embed}\left({n}_{t}\right)$$\end{document} In Eq. (7), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}\in\:{\mathbb{R}}^{d}$$\end{document} denotes the embedding vector for news item \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document}, where d is the embedding dimension. The embedding function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{f}_{embed}$$\end{document} can be implemented either through an embedding matrix lookup or constructed via deep neural networks, facilitating the transformation of raw features into compact vector forms30. After embedding each news article within the user behavior sequence, a user behavior sequence representation matrix is derived, as expressed in Eq. (8):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{E}=\left[{e}_{1},{e}_{2},\cdots\:,{e}_{L}\right],E\in\:{R}^{L\times\:d}$$\end{document} Building upon the user behavior sequence embedding, temporal dynamics and higher-order interactions within the sequence are further modeled. User behavior sequences inherently exhibit temporal dependencies, where recent actions more accurately reflect current user interests compared to long-term behaviors31. To capture this effect, a temporal decay weighting mechanism is integrated with FCNN architectures. In this mechanism, the influence of past behaviors gradually diminishes over time, assigning greater importance to more recent actions. Given a behavior sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\left\{{n}_{1},{n}_{2},\cdots\:,{n}_{L}\right\}$$\end{document}, the decay weight for the -th behavior is defined as Eq. (9):9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{t}=\mathrm{e}\mathrm{x}\mathrm{p}(-\lambda\:\cdot\:\left(L-t\right))$$\end{document} In Eq. (9), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:>0$$\end{document} represents the temporal decay coefficient, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{t}$$\end{document} denotes the weight assigned to the t-th behavior. This weight is applied to the embedding matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{E}$$\end{document} to derive a weighted sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}^{{\prime\:}}$$\end{document}, as shown in Eq. (10):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}^{{\prime\:}}=\mathrm{W}\odot\:\mathrm{E}$$\end{document} In Eq. (10), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{W}=\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\left(\left[{w}_{1},{w}_{2},\cdots\:,{w}_{L}\right]\right)$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\odot\:$$\end{document} indicates element-wise multiplication. To capture higher-order features embedded within the weighted behavior sequence, nonlinear transformations are applied to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}^{{\prime\:}}$$\end{document} through multiple fully connected layers32, as shown in Eq. (11):11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}^{\left(l\right)}=f({W}^{\left(l\right)}{H}^{\left(l-1\right)}+{b}^{\left(l\right)})$$\end{document} In Eq. (11), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}^{\left(l\right)}$$\end{document} represents the output of the -th layer, with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}^{\left(0\right)}={E}^{{\prime\:}}$$\end{document} as the initial input. The parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}^{\left(l\right)}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}^{\left(l\right)}$$\end{document} are the layer, while the activation function f is typically implemented as ReLU in this context. Following multiple nonlinear transformations, the final user sequence representation is achieved, as indicated in Eq. (12):12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{u}=\mathrm{F}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\left({H}^{\left(D\right)}\right)$$\end{document} In Eq. (12), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{u}$$\end{document} serves as a fixed-dimensional, higher-order feature vector that forms the basis for subsequent multi-task optimization, and D denotes the number of network layers. To enhance the depth of the user behavior sequence representation, a multimodal feature fusion mechanism is introduced33. Each news article \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document} is characterized by multiple feature types: (1) Title and content embedding: Semantic vectors are derived using pretrained models such as BERT. (2) Category embedding: News categories are first one-hot encoded and then mapped into a dense embedding space. (3) Time information embedding: The timestamp of the user’s click is recorded and converted into a periodic time embedding vector. By concatenating these feature vectors, a comprehensive embedding vector is formed, as expressed in Eq. (13):13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}=\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}({e}_{t}^{title},{e}_{t}^{category},{e}_{t}^{time})$$\end{document} In this context, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}^{title}$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}^{category}$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}^{time}$$\end{document} represent the embedding vectors of the news title, category, and time features, respectively. This multimodal fusion strategy enriches the semantic richness and precision of the news embedding representation. Multi-task learning framework To elevate recommendation system performance, a multi-task learning framework is proposed, which concurrently optimizes user click intention prediction and user interest distribution learning. In this framework, a shared feature extraction module forms the basis for two parallel tasks, jointly enhancing both the precision of click prediction and the diversity of user interests. This dual-task approach fosters greater accuracy in recommendation outcomes and improves the coverage of results and adaptability in cold-start scenarios. Cold-start users are defined as those who clicked no more than five news items during the data collection window, following common practices in news recommendation research (e.g., Wu et al., 20231. This threshold balances representativeness and noise by avoiding overly strict (≤ 3) or overly lenient (≤ 10) criteria. A time-sensitive data partitioning strategy is adopted to reflect real-world recommendation contexts. For each user, the first four click events in chronological order are used as historical input to the model, while the fifth click event serves as the prediction target for validation or testing. The full dataset is partitioned into training, validation, and test sets in a 7:2:1 ratio. Importantly, cold-start users are included exclusively in the test set to ensure the integrity and realism of evaluation. This partitioning strategy preserves the unique behavioral characteristics of cold-start users while avoiding data leakage through strict temporal isolation. During model training, only complete behavior sequences from non-cold-start users are used for updating model parameters. The limited interaction data from cold-start users is withheld until the final evaluation phase. This approach closely simulates real-world deployment, where models must generate recommendations for users with sparse historical data without relying on prior exposure to their behavior. The design of the multi-task learning framework is grounded in the theoretical principle that task relatedness promotes generalization. Within this framework, the click prediction task—focused on modeling explicit user feedback—and the interest distribution task—aimed at capturing implicit preference patterns—jointly utilize the shared embedding representations derived from user behavior sequences. This shared structure encourages the extraction of common features such as latent semantic preferences, while the task-specific layers are tailored to meet the unique demands of each objective: immediate responsiveness for click prediction and long-term consistency for interest distribution modeling. Empirical findings by Hao et al. (2023) support the efficacy of this approach, demonstrating that multi-task learning can effectively mitigate data sparsity through parameter sharing34. In cold-start scenarios, where user behavior data is sparse, the auxiliary supervision provided by the interest distribution task acts as a regularizer, structuring the embedding space and reducing the risk of overfitting. The multi-task framework is built around a shared feature extraction module that provides a common input for multiple objective tasks35. This structure enables collaborative optimization, resulting in an overall enhancement in model performance. A section of the multi-task learning framework code is illustrated in Fig. 2. Fig. 2Multi-Task learning code section. Multi-Task learning code section. Figure 2 depicts the shared feature layer, responsible for extracting high-order features from the user behavior sequence. These features serve as common inputs for the two parallel tasks within the multi-task learning framework. The first task, click intention prediction, estimates the probability that a user will click on a particular news article, optimizing recommendation accuracy. The second task, interest distribution learning, forecasts the user’s interest distribution across various domains, thereby enhancing recommendation diversity. After processing the user behavior sequence, the model passes the high-order features from the shared layer to each task module. Each module independently optimizes its respective objective function, with overall optimization achieved through a weighted joint loss. The objective of the user click intention prediction task is to determine the likelihood of a user clicking on a specific news article. This task is framed as a binary classification problem, with the output, the click probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{y}\in\:\left[\mathrm{0,1}\right]$$\end{document}, computed using Eq. (14):14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{y}={\upsigma\:}({W}_{click}^{T}{h}_{u}+{b}_{click})$$\end{document} In Eq. (14), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{u}$$\end{document} represents the output from the shared feature layer, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{click}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{click}$$\end{document} are learnable parameters, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\upsigma\:}(\bullet\:)$$\end{document} is the Sigmoid activation function. The optimization of this task enhances the accuracy of recommendation outcomes, allowing for a more precise capture of user click preferences36. Recommendation systems should accurately forecast user clicks while ensuring content diversity to maximize user satisfaction. To achieve this, a user interest distribution learning task is introduced, designed to capture the distribution characteristics of user interests across various domains. The user interest distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p\in\:{\mathbb{R}}^{K}$$\end{document} is formulated as shown in Eq. (15):15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}({W}_{interest}^{T}{h}_{u}+{b}_{interest})$$\end{document} In Eq. (15), K represents the number of interest categories, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{W}_{interest}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{b}_{interest}$$\end{document} denote the parameter matrix and bias vector for interest distribution prediction, and the Softmax function normalizes the output into a probability distribution. The objective of the interest distribution learning task is to minimize the error between the true user interest distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{true}$$\end{document} and the predicted distribution p. Optimizing this module enhances the recommendation system’s content diversity, ensuring that the results comprehensively reflect the user’s broader interests37. To synchronize the optimization of both tasks, a weighted joint loss function is utilized, combining the goals of click intention prediction and interest distribution learning38. This joint optimization enables the model to simultaneously refine the accuracy and diversity of recommendations, leading to a marked improvement in overall performance. Cold-start users are defined based on the sparsity of their historical behavior data. Specifically, users who clicked no more than five news items within the data collection window are classified as cold-start users. To ensure fairness and rigor in evaluation, a strict data partitioning and isolation strategy is adopted. All identified cold-start users are assigned exclusively to the test set, meaning they remain completely unseen during the training and validation stages. Model parameter updates rely solely on the behavior sequences of non–cold-start users with sufficient data. At the final testing stage, the model uses the limited historical click behaviors of cold-start users (no more than five clicks) as input to evaluate recommendation performance under real cold-start scenarios. This design eliminates the risk of data leakage and guarantees the validity and reproducibility of the experimental results. To further support academic exchange and research verification, the related code, data-splitting scripts, and experimental configurations are made available on the hosting platform (https://github.com/C3R8U/MT-FCNN-for-Personalized-News-Recommendation). The main mathematical symbols involved in the MT-FCNN model are uniformly defined in Table 1. Table 1Symbol definition.SymbolMathematical FormMeaningUser b Represents an individual userNews item \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document} The t-th clicked news itemSequence length L Length of the user historical behavior sequenceTime index t Time step in the sequenceEmbedding vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}\in\:{\mathbb{R}}^{d}$$\end{document} Feature embedding of the t-th news itemEmbedding matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{E}\in\:{\mathbb{R}}^{L\times\:d}$$\end{document} Embedding matrix of the user behavior sequenceWeight coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{t}$$\end{document} Temporal decay weightDecay parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} Temporal decay coefficientWeight matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{W}=\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\left(\left[{w}_{1},{w}_{2},\cdots\:,{w}_{L}\right]\right)$$\end{document} Diagonal matrix of temporal weightsLayer output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}^{\left(l\right)}$$\end{document} Output of the l-th FCNN layerNetwork depth D Number of FCNN layersUser representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{u}$$\end{document} Final user behavior representationClick prediction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{y}$$\end{document} Predicted click probabilityClick label y Ground-truth click labelInterest distribution p Predicted interest distributionTrue distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{p}$$\end{document} Ground-truth interest distributionClick loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{click}$$\end{document} Loss for click prediction taskInterest loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{interest}$$\end{document} Loss for interest distribution taskTotal loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{total}$$\end{document} Multi-task joint lossTask weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} Balancing coefficient for multi-task learningHidden dimension k Dimensionality of FCNN hidden layersEmbedding dimension d Feature embedding dimensionNumber of interest classes K Number of interest categories Table 1Symbol definition.SymbolMathematical FormMeaningUser b Represents an individual userNews item \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{n}_{t}$$\end{document} The t-th clicked news itemSequence length L Length of the user historical behavior sequenceTime index t Time step in the sequenceEmbedding vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{e}_{t}\in\:{\mathbb{R}}^{d}$$\end{document} Feature embedding of the t-th news itemEmbedding matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{E}\in\:{\mathbb{R}}^{L\times\:d}$$\end{document} Embedding matrix of the user behavior sequenceWeight coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{t}$$\end{document} Temporal decay weightDecay parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\lambda\:$$\end{document} Temporal decay coefficientWeight matrix \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\mathrm{W}=\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{g}\left(\left[{w}_{1},{w}_{2},\cdots\:,{w}_{L}\right]\right)$$\end{document} Diagonal matrix of temporal weightsLayer output \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{H}^{\left(l\right)}$$\end{document} Output of the l-th FCNN layerNetwork depth D Number of FCNN layersUser representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{u}$$\end{document} Final user behavior representationClick prediction \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{y}$$\end{document} Predicted click probabilityClick label y Ground-truth click labelInterest distribution p Predicted interest distributionTrue distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{p}$$\end{document} Ground-truth interest distributionClick loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{click}$$\end{document} Loss for click prediction taskInterest loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{interest}$$\end{document} Loss for interest distribution taskTotal loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mathcal{L}}_{total}$$\end{document} Multi-task joint lossTask weight \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} Balancing coefficient for multi-task learningHidden dimension k Dimensionality of FCNN hidden layersEmbedding dimension d Feature embedding dimensionNumber of interest classes K Number of interest categories Symbol definition. The overall training and inference procedure of MT-FCNN is further abstracted into a standard algorithmic workflow. This procedure covers user behavior sequence embedding, FCNN-based feature extraction, multi-task joint optimization, and parameter updating, thereby providing a systematic description of the model training mechanism, as shown in Table 2. Table 2Pseudocode of the MT-FCNN Algorithm.StepAlgorithm DescriptionInputUser behavior sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{u}=\{{n}_{1},{n}_{2},\dots\:,{n}_{L}\}$$\end{document}, news feature setOutputClick probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{y}$$\end{document}, interest distribution p1Initialize embedding layer parameters, FCNN shared layers, and task-specific parameters2Perform feature embedding on user behavior sequence to obtain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:E=\{{e}_{1},{e}_{2},\dots\:,{e}_{L}\}$$\end{document}3Compute temporal decay weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{t}$$\end{document} and construct weight matrix W4Obtain weighted embedding representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}^{{\prime\:}}=WE$$\end{document}5Feed (E’) into FCNN shared network for multi-layer nonlinear transformation6Obtain final user representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{u}=\mathrm{F}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\left({H}^{\left(D\right)}\right)$$\end{document}7Click prediction task: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{y}=\sigma\:({W}_{click}^{T}{h}_{u}+{b}_{click})$$\end{document}8Interest distribution task: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}({W}_{interest}^{T}{h}_{u}+{b}_{interest})$$\end{document}9Compute click loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{click}$$\end{document}and interest loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{interest}$$\end{document}10Construct multi-task joint loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{total}=\alpha\:{L}_{click}+(1-\alpha\:){L}_{interest}$$\end{document}11Update model parameters using Adam optimizer12Repeat steps 2–11 until convergenceOutputTrained MT-FCNN model Table 2Pseudocode of the MT-FCNN Algorithm.StepAlgorithm DescriptionInputUser behavior sequence \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{S}_{u}=\{{n}_{1},{n}_{2},\dots\:,{n}_{L}\}$$\end{document}, news feature setOutputClick probability \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{y}$$\end{document}, interest distribution p1Initialize embedding layer parameters, FCNN shared layers, and task-specific parameters2Perform feature embedding on user behavior sequence to obtain \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:E=\{{e}_{1},{e}_{2},\dots\:,{e}_{L}\}$$\end{document}3Compute temporal decay weights \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{w}_{t}$$\end{document} and construct weight matrix W4Obtain weighted embedding representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{E}^{{\prime\:}}=WE$$\end{document}5Feed (E’) into FCNN shared network for multi-layer nonlinear transformation6Obtain final user representation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{h}_{u}=\mathrm{F}\mathrm{l}\mathrm{a}\mathrm{t}\mathrm{t}\mathrm{e}\mathrm{n}\left({H}^{\left(D\right)}\right)$$\end{document}7Click prediction task: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\widehat{y}=\sigma\:({W}_{click}^{T}{h}_{u}+{b}_{click})$$\end{document}8Interest distribution task: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:p=\mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}({W}_{interest}^{T}{h}_{u}+{b}_{interest})$$\end{document}9Compute click loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{click}$$\end{document}and interest loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{interest}$$\end{document}10Construct multi-task joint loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{L}_{total}=\alpha\:{L}_{click}+(1-\alpha\:){L}_{interest}$$\end{document}11Update model parameters using Adam optimizer12Repeat steps 2–11 until convergenceOutputTrained MT-FCNN model Pseudocode of the MT-FCNN Algorithm. Training and optimization of the model To effectively train the MT-FCNN model and optimize its performance in recommendation tasks, a detailed design of the training process, optimization algorithm, and parameter adjustment strategy is implemented. The training data utilized in this study is sourced from the publicly available MIND (Microsoft News Dataset), which randomly selects 1 million users who have interacted with at least five news articles. User privacy is preserved through secure hashing, anonymizing each user’s ID. The Adam optimizer is employed to optimize the objective function, as it is efficient for large-scale deep learning models and demonstrates rapid convergence39. The parameter update rule is expressed in Eq. (16):16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}={\theta\:}_{t}-\eta\:\frac{{\widehat{m}}_{t}}{\sqrt{{\widehat{v}}_{t}}+ϵ}$$\end{document} In Eq. (16), \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document} represents the learning rate, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{m}}_{t}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\widehat{v}}_{t}$$\end{document} are the first-order and second-order moment estimates, respectively, and ϵ is a small constant introduced to avoid division by zero. In model training, hyperparameter configuration plays a crucial role in optimizing performance. The settings of these hyperparameters in this study are derived from experimental outcomes and relevant literature. Table 3 presents a comprehensive overview of the hyperparameter settings and adjustments made throughout the training process. Table 3Hyperparameter Settings and Adjustments.HyperparameterDescriptionSetting ValueLearning Rate ηThe initial learning rate is set to 10− 3, with a decay strategy applied when validation performance stagnates. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{-3}$$\end{document} Learning Rate DecayThe learning rate is decayed by a factor of 0.1 when no improvement in validation performance is observed.0.1 times decayBatch Size BBBThe size of mini-batches used in gradient descent for efficient training.128Task Weight αThe weight assigned to each task in the multi-task learning framework, determined via grid search.Grid search selected: 0.5, with values tested: 0.3, 0.5,Dropout RateA technique used to prevent overfitting by randomly deactivating units during training.0.5L2 RegularizationA regularization method that applies weight decay to prevent overfitting. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{-4}$$\end{document} Table 3Hyperparameter Settings and Adjustments.HyperparameterDescriptionSetting ValueLearning Rate ηThe initial learning rate is set to 10− 3, with a decay strategy applied when validation performance stagnates. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{-3}$$\end{document} Learning Rate DecayThe learning rate is decayed by a factor of 0.1 when no improvement in validation performance is observed.0.1 times decayBatch Size BBBThe size of mini-batches used in gradient descent for efficient training.128Task Weight αThe weight assigned to each task in the multi-task learning framework, determined via grid search.Grid search selected: 0.5, with values tested: 0.3, 0.5,Dropout RateA technique used to prevent overfitting by randomly deactivating units during training.0.5L2 RegularizationA regularization method that applies weight decay to prevent overfitting. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{10}^{-4}$$\end{document} Hyperparameter Settings and Adjustments. Results and discussions Model performance evaluation Figure 3 presents a performance comparison of various models on the MIND dataset. The MT-FCNN model significantly outperforms all alternatives across every evaluated metric. The baseline models include User-based Collaborative Filtering (User-CF), which relies on the user–news interaction matrix; TF-IDF Content-Based Recommendation (CBR-TFIDF), which is based on textual feature similarity; GRU4Rec, a sequential modeling approach; and a Transformer-based news recommendation model that leverages the self-attention mechanism. Notably, improvements of 12.4% in AUC and 10.7% in NDCG are observed, alongside an 8.9% increase in CTR when compared to conventional approaches. Specifically, MT-FCNN achieves an AUC of 94.5%, which surpasses the User-CF model’s 75.2% by 19.3% points, the CBR-TFIDF model’s 78.3% by 16.2% points, the GRU4Rec model’s 82.5% by 12% points, and the Transformer-based model’s 84.0748% by 10.4252% points. In terms of NDCG@5 and NDCG@10, MT-FCNN secures 61% and 65.7%, respectively, which significantly exceed the performance of other models. The CTR metric also favors MT-FCNN, which leads with a 9.1% improvement, outperforming the User-CF model by 2.6% points, the CBR-TFIDF model by 2% points, the GRU4Rec model by 1.3% points, and the Transformer-based model by 0.7436% points. These results highlight the MT-FCNN model’s superior performance in recommendation accuracy and user CTR. Fig. 3Performance Comparison of Different Models in the Cold-Start User Scenario. Performance Comparison of Different Models in the Cold-Start User Scenario. In the cold-start user scenario, the MT-FCNN model demonstrates superior recommendation accuracy. MT-FCNN achieves a Precision@5 of 47.2%, compared to 30.2% for User-CF, 35.4% for CBR-TFIDF, 40% for GRU4Rec, and 42.4% for the Transformer-based model. For Recall@5, MT-FCNN leads with a score of 23.6%, surpassing the User-CF model by 56.3%, the CBR-TFIDF model by 33.3%, the GRU4Rec model by 18%, and the Transformer-based model by 10.4%. MT-FCNN achieves an F1-score@5 of 34.2%, which is 56.2% higher than that of User-CF. These results highlight MT-FCNN’s notable advantage in recommendation accuracy for cold-start users and its ability to enhance recommendation performance for new users Fig. 4. Fig. 4Training efficiency comparison of different models. Training efficiency comparison of different models. Training efficiency comparisons across various models are illustrated in Fig. 5. The MT-FCNN model strikes a favorable balance in training efficiency. Despite requiring 4 h of training time—slightly longer than the User-CF and CBR-TFIDF models—it outperforms the GRU4Rec and Transformer-based models, which take 5 and 6 h, respectively. In terms of convergence epochs, MT-FCNN reaches convergence in 20 epochs, fewer than the 30 epochs required by GRU4Rec and the 25 epochs by the Transformer-based model, indicating a faster convergence rate. MT-FCNN also has a more efficient parameter count, with 4.5 million parameters compared to GRU4Rec’s 5 million and the Transformer-based model’s 8 million. Regarding memory usage, MT-FCNN demands only 1.5 GB, significantly lower than the 2 GB required by GRU4Rec and the 3 GB by the Transformer-based model, demonstrating superior memory efficiency. In sum, MT-FCNN delivers high performance while ensuring optimal resource utilization and training efficiency. Fig. 5Training efficiency comparison of different models. Training efficiency comparison of different models. Ablation study on model components The results of the ablation study on key components of the MT-FCNN model are depicted in Fig. 6, providing a detailed assessment of the individual contributions to the model’s performance. All experiments are repeated 10 times under identical data splits and parameter settings, and the results are reported in terms of mean and standard deviation to ensure a more reliable evaluation of stability and reproducibility. The removal of the multi-task learning module results in a significant performance reduction, with AUC decreasing to 90.2%, NDCG@5 declining to 56.3%, and CTR dropping to 8.5%. These results highlight the pivotal role of multi-task learning in simultaneously optimizing recommendation accuracy and diversity. Excluding the user behavior sequence embedding leads to further performance degradation, with AUC reduced to 88.7%, NDCG@5 to 54.8%, and CTR to 8.2%. This observation emphasizes the embedding’s necessity for modeling the temporal dynamics and evolving patterns in user interests. The removal of the FCNN leads to a substantial degradation in performance, with AUC decreasing to 85%, NDCG@5 dropping to 50.5%, and CTR falling to 7.9%. This sharp decline indicates the crucial role of fully connected layers in capturing higher-order feature interactions and strengthening representation learning. The ablation results further demonstrate that the multi-task learning mechanism, behavior sequence embeddings, and FCNN architecture exhibit strong synergistic effects, jointly forming the fundamental basis for the performance gains achieved by the MT-FCNN model. Fig. 6Results of the ablation study on model components. Results of the ablation study on model components. The MT-FCNN model demonstrates significant superiority in metrics related to recommendation diversity, coverage, and novelty. The diversity index for MT-FCNN reaches 0.65, a notable improvement compared to other models. Specifically, User-CF achieves 0.45, CBR-TFIDF records 0.5, GRU4Rec attains 0.55, and Transformer-based models achieve 0.58. This indicates that MT-FCNN provides a broader range of content. In terms of recommendation coverage, the MT-FCNN model achieves a coverage rate of 55%, significantly outperforming User-CF (35%), CBR-TFIDF (40.5%), GRU4Rec (45.8%), and Transformer-based models (48.2%). This demonstrates an enhanced capability to cover diverse domains and topics, extending the model’s adaptability across various user interest distributions. For novelty, MT-FCNN delivers a performance of 52.5%, further surpassing the benchmarks set by other models. This metric highlights the model’s strength in introducing fresh and less familiar content, enriching the user experience with innovative recommendations. These results collectively affirm the MT-FCNN model’s robust capacity to generate diverse, expansive, and novel recommendations. Such attributes help satisfy diverse user preferences and improve overall satisfaction Fig. 7. Fig. 7Diversity metrics of different models. Diversity metrics of different models. Figure 8 illustrates the user satisfaction scores across different models. The MT-FCNN model demonstrates exceptional performance in user satisfaction metrics. MT-FCNN achieves the highest satisfaction score of 8.3, significantly surpassing User-CF (6.5), CBR-TFIDF (6.8), GRU4Rec (7.2), and Transformer-based models (7.5). This result underscores the model’s ability to deliver a highly engaging and user-centric recommendation experience. In terms of willingness to continue system usage, the MT-FCNN model achieves an 82.5% rate, markedly higher than User-CF (60%), CBR-TFIDF (62.5%), GRU4Rec (70%), and Transformer-based models (75%). This metric reflects the sustained user engagement and trust cultivated by the model’s recommendations. For recommendation willingness, the MT-FCNN model secures a 78% rate, exceeding the levels of User-CF (55%), CBR-TFIDF (57.5%), GRU4Rec (65%), and Transformer-based models (70%). Such a result highlights the model’s capacity to generate suggestions that users are inclined to share with others, indicative of its relevance and perceived value. These findings indicate that MT-FCNN not only excels in quantitative metrics but also gains strong user approval, highlighting its practical applicability. Fig. 8User satisfaction scores across different models. User satisfaction scores across different models. Empirical studies have shown that multi-task learning can effectively mitigate data sparsity through parameter sharing. In cold-start scenarios, where user behavior data is extremely sparse, the auxiliary supervision signal provided by the interest distribution task plays a crucial regularization role. It constrains the structure of the shared embedding space and guides the model to learn more generalizable features, thereby reducing the risk of overfitting. This mechanism is fundamentally different from simple data augmentation, which aims to increase the volume of data, whereas multi-task learning improves representational quality by introducing structured information from related tasks. This study conducts systematic ablation experiments to quantitatively evaluate the mutual influence between the click prediction task and the interest distribution modeling task. Specifically, based on the original multi-task framework, two single-task model variants are constructed: (1) the click prediction-only model (α = 1), which completely removes the interest distribution learning task; (2) the interest distribution-only model (α = 0), which optimizes only the interest distribution prediction. Other hyperparameters remain consistent with the original model, and each variant is tested using the same data splits with five repeated experiments to assess stability. The experimental results, shown in Fig. 9, indicate that the complete multi-task model significantly outperforms the click prediction-only model in terms of AUC (94.5% vs. 93.1%) and NDCG@5 (61.0% vs. 58.7%) (p < 0.01), demonstrating that introducing the interest distribution task indirectly improves the accuracy of click prediction by capturing the diversity of user interests. Meanwhile, although the interest distribution-only model achieves coverage close to the full model (53.8%), its AUC (87.6%) and NDCG@5 (52.4%) drop considerably, indicating that modeling interest distribution alone lacks the ability to capture users’ immediate intentions. Further analysis shows that removing the interest distribution task leads to a 27.3% decrease in category coverage (from 55.0% to 40.2%), causing insufficient exposure of long-tail news. Conversely, removing the click prediction task results in relatively high coverage but a significant decline in relevance to users’ real-time interests (NDCG@5 drops by 14.0%). This phenomenon confirms the complementarity of the two tasks within the multi-task framework: the click prediction task enhances recommendation relevance by accurately capturing immediate user behaviors, while the interest distribution task broadens recommendation diversity by mining latent interest dimensions. Their joint optimization balances accuracy and diversity objectives and strengthens the model’s generalization ability for sparse behavior data, such as in cold-start scenarios. For example, within the cold-start user subset, the full model’s coverage (48.5%) improves by 51.1% compared to the click prediction-only model (32.1%), further validating the adaptability of multi-task collaboration in complex recommendation environments. Overall, the full MT-FCNN consistently outperforms the single-task CTR-only variant. This demonstrates that the introduction of the interest distribution task avoids negative transfer while enhancing overall performance through positive knowledge transfer, highlighting the complementarity between tasks. Fig. 9Ablation experiment results on multi-task collaboration. Ablation experiment results on multi-task collaboration. Figure 9 presents the ablation experiment results on multi-task collaboration, systematically investigating the impact of the task weight coefficient α on model performance and validating the robustness of the multi-task framework design. A grid search was conducted over α ∈ {0.3, 0.4, 0.5, 0.6, 0.7}, with five repeated experiments for each α value to assess result stability. The outcomes are summarized in Table 4. When α = 0.5, the model achieves optimal performance, with an AUC of 94.5% and an NDCG@5 of 61.0%, while also maintaining a high diversity index of 0.65. This suggests that a balanced weighting between the click prediction task (α = 0.5) and the interest distribution task (1 - α = 0.5) effectively reconciles the objectives of recommendation accuracy and diversity. When α < 0.5 (e.g., α = 0.3), the increased emphasis on the interest distribution task enhances coverage (62.1%), but the AUC drops by 1.7% points, indicating that overemphasizing diversity may weaken the model’s ability to capture immediate user intent. Conversely, when α > 0.5 (e.g., α = 0.7), the click prediction task dominates, resulting in a 1.5% point decrease in AUC and a notable drop in coverage to 49.3%, highlighting the limitations of single-task optimization. Further analysis in the cold-start scenario shows that at α = 0.5, the AUC for cold-start users (47.2%) improves by 7.8% and 4.7% compared to α = 0.3 (43.8%) and α = 0.7 (45.1%), respectively. This demonstrates that balanced task weighting enhances generalization to sparse data by leveraging shared feature learning. These findings align with the multi-task learning principle that “task relatedness promotes generalization.” The α = 0.5 configuration is supported by both theoretical rationale and empirical results, establishing a solid foundation for the model’s robustness. Table 4Model performance comparison under different α values (Mean ± Standard Deviation).\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{\alpha\:}$$\end{document} ValueAUC (%)NDCG@5 (%)Coverage@5 (%)Diversity Index0.392.8 ± 0.458.3 ± 0.362.1 ± 0.50.68 ± 0.020.493.5 ± 0.359.1 ± 0.258.4 ± 0.40.63 ± 0.030.5 94.5 ± 0.3 61.0 ± 0.2 55.0 ± 0.4 0.65 ± 0.02 0.693.9 ± 0.460.2 ± 0.351.7 ± 0.60.59 ± 0.040.793.0 ± 0.559.5 ± 0.449.3 ± 0.70.54 ± 0.05 Table 4Model performance comparison under different α values (Mean ± Standard Deviation).\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\boldsymbol{\alpha\:}$$\end{document} ValueAUC (%)NDCG@5 (%)Coverage@5 (%)Diversity Index0.392.8 ± 0.458.3 ± 0.362.1 ± 0.50.68 ± 0.020.493.5 ± 0.359.1 ± 0.258.4 ± 0.40.63 ± 0.030.5 94.5 ± 0.3 61.0 ± 0.2 55.0 ± 0.4 0.65 ± 0.02 0.693.9 ± 0.460.2 ± 0.351.7 ± 0.60.59 ± 0.040.793.0 ± 0.559.5 ± 0.449.3 ± 0.70.54 ± 0.05 Model performance comparison under different α values (Mean ± Standard Deviation). This study introduces the Shapley Additive Explanations (SHAP) method to quantify the contributions of features such as title embedding and category embedding to click prediction and employs visual analysis to uncover the model’s decision-making logic. A random sample of 1,000 users from the test set, including 300 cold-start users, was selected for evaluation. The mean SHAP values of each feature dimension in the users’ historical behavior sequences were computed, and the results are presented in Fig. 10. The analysis reveals that title embedding, with a mean SHAP value of 0.45, is the most influential feature in click prediction, accounting for 42.1% of the total contribution. This influence is substantially greater than that of category embedding (0.28) and time embedding (0.15). As shown in Fig. 10, the SHAP value distribution for title embedding is right-skewed, indicating a stronger positive impact on predicting clicks (i.e., positive samples). Specifically, when the title embedding semantically aligns with a user’s historical interests (e.g., a user interested in “technology” receiving headlines related to “artificial intelligence”), the SHAP value can exceed 0.6. In contrast, when the title’s semantics diverge from the user’s interest (e.g., a “sports” user receiving a “finance” headline), the SHAP value typically falls below 0.2. This pattern confirms the model’s ability to accurately capture real-time user intention through title semantics. Notably, the contribution of title embeddings for cold-start users (mean SHAP value = 0.38) is slightly lower than that for regular users but remains significantly higher than other features, such as category embedding (0.31). This suggests that the model can infer latent interests from title content even when behavioral data is limited. For instance, a cold-start user with only three recorded clicks, who had not previously interacted with medical news, was recommended a health-related article due to the presence of the keyword “vaccine development” in the title. The corresponding SHAP value reached 0.41, demonstrating that semantic information in titles can effectively compensate for sparse behavioral data. Moreover, the increased contribution of category embedding in cold-start scenarios (0.31 vs. 0.28 in regular users) reflects the model’s reliance on explicit categorical information to generalize interests under data scarcity. These findings validate the design of MT-FCNN from an interpretability perspective and offer practical insights for news platforms: enhancing the semantic clarity and distinctiveness of titles can further improve the relevance of recommendations. Fig. 10Feature contribution analysis via SHAP values. Feature contribution analysis via SHAP values. To enhance the qualitative evaluation of model interpretability, a case study of representative users is conducted. For an active user whose historical click behavior is mainly concentrated on technology-related news such as “artificial intelligence” and “machine learning,” the model prioritizes recommending content related to “recent advances in deep learning.” The SHAP analysis indicates that title semantics contribute significantly more than category and temporal features, thereby providing an explanation for the generated recommendation results. For a cold-start user case, the historical behavior consists of only a small number of entertainment-related clicks. Nevertheless, the model successfully recommends relevant content based on keywords such as “movie” and “box office” in the news titles. This decision process is consistent with the high-contribution features identified by SHAP, validating the model’s interpretable decision-making capability under sparse data conditions. Overall, these case studies provide further evidence at the individual level for the rationality of the model’s decision logic and are consistent with the global SHAP-based statistical results, thereby offering mutual validation. This study quantifies the specific impact of the interest distribution task on recommendation diversity by comparing the complete MT-FCNN model with a variant in which the interest distribution task is removed. Experiments were conducted on the test set to compute Coverage@5 (the proportion of news categories represented in the top-5 recommendations) and Gini@5 (a measure of the concentration or inequality in the recommendation distribution). The results are illustrated in Fig. 11. The findings reveal a substantial reduction in diversity when the interest distribution task is omitted. Specifically, Coverage@5 decreases from 55.0% to 40.5%, representing a 26.4% decline, while the Gini coefficient increases from 0.25 to 0.38—a 52.0% relative increase. This indicates that, without the interest distribution task, the model’s recommendations become significantly more concentrated in popular categories such as “sports” and “entertainment,” at the expense of long-tail categories like “technology” and “health.” For example, the complete model covers 32 news subcategories in the test set, whereas the variant without the interest distribution task covers only 24. Furthermore, the proportion of recommendations accounted for by the top three most frequently recommended categories rises from 47.8% to 65.3%. These results underscore the essential role of the interest distribution task in enhancing category-level diversity by guiding the model to explore a wider range of user interests beyond immediate behavioral cues. Fig. 11Interest distribution task and its impact on recommendation diversity. Interest distribution task and its impact on recommendation diversity. Repeated experiments and hypothesis testing results Under the same experimental settings, ten independent repeated experiments are conducted to verify the stability and statistical significance of the proposed MT-FCNN model and four baseline methods, including User-CF, CBR-TFIDF, GRU4Rec, and a Transformer-based news recommendation model. In each run, model parameters were re-initialized, and the dataset was randomly partitioned to eliminate the influence of randomness. The experimental results are summarized in Table 5, with key evaluation metrics—AUC, NDCG@5, and CTR—reported as the mean ± standard deviation. To assess the statistical significance of performance differences, two-tailed paired t-tests were performed between MT-FCNN and each baseline model, and the corresponding p-values are reported. Table 5Statistical significance analysis of 10 experimental runs.ModelAUC (%)NDCG@5 (%)CTR (%)t-value and p-value vs. MT-FCNN (AUC)t-value and p-value vs. MT-FCNN (NDCG@5)t-value and p-value vs. MT-FCNN (CTR)MT-FCNN94.5 ± 0.361.0 ± 0.29.1 ± 0.1---User-CF75.2 ± 1.130.2 ± 0.86.5 ± 0.3t = 6.82, p < 0.001t = 7.11, p < 0.001t = 6.95, p < 0.001CBR-TFIDF78.3 ± 0.935.4 ± 0.67.1 ± 0.2t = 6.41, p < 0.001t = 6.78, p < 0.001t = 6.52, p < 0.001GRU4Rec82.5 ± 0.640.0 ± 0.47.8 ± 0.2t = 5.93, p < 0.001t = 6.21, p < 0.001t = 6.05, p < 0.001Transformer84.1 ± 0.542.4 ± 0.38.4 ± 0.1t = 5.27, p < 0.001t = 5.63, p < 0.001t = 5.41, p < 0.001 Table 5Statistical significance analysis of 10 experimental runs.ModelAUC (%)NDCG@5 (%)CTR (%)t-value and p-value vs. MT-FCNN (AUC)t-value and p-value vs. MT-FCNN (NDCG@5)t-value and p-value vs. MT-FCNN (CTR)MT-FCNN94.5 ± 0.361.0 ± 0.29.1 ± 0.1---User-CF75.2 ± 1.130.2 ± 0.86.5 ± 0.3t = 6.82, p < 0.001t = 7.11, p < 0.001t = 6.95, p < 0.001CBR-TFIDF78.3 ± 0.935.4 ± 0.67.1 ± 0.2t = 6.41, p < 0.001t = 6.78, p < 0.001t = 6.52, p < 0.001GRU4Rec82.5 ± 0.640.0 ± 0.47.8 ± 0.2t = 5.93, p < 0.001t = 6.21, p < 0.001t = 6.05, p < 0.001Transformer84.1 ± 0.542.4 ± 0.38.4 ± 0.1t = 5.27, p < 0.001t = 5.63, p < 0.001t = 5.41, p < 0.001 Statistical significance analysis of 10 experimental runs. As shown in Table 5, MT-FCNN significantly outperforms all baseline models across AUC, NDCG@5, and CTR metrics (p < 0.001). For example, the mean AUC of MT-FCNN (94.5%) surpasses that of the Transformer model (84.1%) by 10.4% points, while its standard deviation (0.3%) is notably lower, indicating greater performance stability. Further analysis confirms that the differences between MT-FCNN and each baseline are statistically highly significant (all p-values < 0.001), rejecting the null hypothesis of no difference with over 99.9% confidence. This provides strong statistical evidence supporting the effectiveness of MT-FCNN’s design, particularly its multi-task learning framework and behavior sequence embeddings. Notably, MT-FCNN achieves an AUC standard deviation of only 0.2% on the cold-start user subset, which is substantially lower than that of GRU4Rec (0.8%) and the Transformer-based model (0.6%), demonstrating superior robustness in sparse data scenarios. In addition, the standard deviation results across ten independent runs further indicate that MT-FCNN consistently exhibits low variance in AUC, NDCG@5, and CTR. This suggests that the model maintains stable performance across different data splits without significant fluctuations. The results also imply that no obvious overfitting behavior is observed during training, and the model is able to maintain consistent recommendation performance under distribution shifts and user behavior sparsity. The experiments were conducted under a unified hardware environment (NVIDIA A100 GPU, batch size = 128, sequence length T = 50) to evaluate each model’s computational efficiency. Table 6 presents the floating-point operations (FLOPs) per forward pass, GPU memory consumption, training time, and model size in terms of the number of parameters. MT-FCNN demonstrates notably lower computational complexity, with only 2.1 × 10⁹ FLOPs—36.2% and 55.3% of the FLOPs required by Transformer (5.8 × 10⁹) and GRU4Rec (3.8 × 10⁹), respectively. This reduction stems from the FCNN architecture, which avoids both the quadratic time complexity (O(T²)) of Transformer’s self-attention and the sequential operations required by GRU layers. In terms of GPU memory usage, MT-FCNN consumes 1.5 GB, which is 50% less than the Transformer model (3.0 GB), making it more feasible for deployment in memory-constrained environments such as edge devices or high-concurrency systems. Regarding training efficiency, MT-FCNN achieves a throughput of 128,000 samples per hour, outperforming Transformer by 50.6% (85,000 samples/hour). This advantage confirms MT-FCNN’s practicality for real-world applications that demand both accuracy and speed. Although the CBR-TFIDF model demonstrates the lowest FLOPs (0.5 × 10⁹) and memory usage (0.8 GB), it also shows inferior predictive performance (AUC = 78.3%), indicating that its efficiency comes at the cost of significantly reduced effectiveness. In contrast, MT-FCNN strikes a more favorable balance between resource efficiency and predictive accuracy (AUC = 94.5%). Sub-experiments conducted under cold-start conditions show that MT-FCNN achieves an inference latency of 2.3 ms per request, a 51.1% reduction compared to Transformer’s 4.7 ms. This low-latency response further validates MT-FCNN’s suitability for time-sensitive, real-time recommendation services. Overall, the results demonstrate that by employing a streamlined, fully connected structure and leveraging multi-task parameter sharing, MT-FCNN achieves substantial reductions in resource consumption while maintaining high recommendation quality, thus offering a scalable and deployable solution for industrial recommendation systems. Table 6Comparison of computational costs across models (mean ± standard deviation).ModelFLOPs (×10⁹)GPU Memory (GB)Training Time (hours)Number of Parameters (Millions)MT-FCNN2.1 ± 0.11.5 ± 0.24.0 ± 0.34.5GRU4Rec3.8 ± 0.22.0 ± 0.35.0 ± 0.55.0Transformer5.8 ± 0.33.0 ± 0.46.0 ± 0.68.0CBR-TFIDF0.5 ± 0.050.8 ± 0.11.5 ± 0.21.2 Table 6Comparison of computational costs across models (mean ± standard deviation).ModelFLOPs (×10⁹)GPU Memory (GB)Training Time (hours)Number of Parameters (Millions)MT-FCNN2.1 ± 0.11.5 ± 0.24.0 ± 0.34.5GRU4Rec3.8 ± 0.22.0 ± 0.35.0 ± 0.55.0Transformer5.8 ± 0.33.0 ± 0.46.0 ± 0.68.0CBR-TFIDF0.5 ± 0.050.8 ± 0.11.5 ± 0.21.2 Comparison of computational costs across models (mean ± standard deviation). To further quantify the computational redundancy of the Transformer architecture and verify the efficiency advantage of the proposed model when processing user behavior sequences of different lengths, a comparison is conducted between MT-FCNN and the Transformer model in terms of FLOPs for single forward propagation at sequence lengths T =20, 50, 100, and 200. The results are presented in Table 7. The Transformer model exhibits a quadratic increase in computational cost as sequence length grew, confirming the complexity introduced by its self-attention mechanism. In contrast, the FLOPs of the MT-FCNN model increase at a much slower, nearly linear rate, demonstrating superior scalability. These results provide strong evidence that the proposed model achieves significant computational efficiency when modeling long user behavior histories, making it more suitable for large-scale, low-latency recommendation system deployment. Table 7FLOPs comparison under different sequence lengths (mean ± standard deviation).Sequence length (T)Transformer FLOPs (×10⁹)MT-FCNN FLOPs (×10⁹)201.1 ± 0.10.9 ± 0.05505.8 ± 0.32.1 ± 0.110023.2 ± 0.74.2 ± 0.220092.5 ± 1.58.5 ± 0.4 Table 7FLOPs comparison under different sequence lengths (mean ± standard deviation).Sequence length (T)Transformer FLOPs (×10⁹)MT-FCNN FLOPs (×10⁹)201.1 ± 0.10.9 ± 0.05505.8 ± 0.32.1 ± 0.110023.2 ± 0.74.2 ± 0.220092.5 ± 1.58.5 ± 0.4 FLOPs comparison under different sequence lengths (mean ± standard deviation). FCNN depth and hyperparameter sensitivity analysis Although traditional RNNs can capture sequential dependencies, their recursive structure is highly prone to gradient vanishing or exploding when processing long sequences. This occurs because gradients are repeatedly multiplied across time steps, leading to exponential decay or growth in their values. In addition, the lack of parallelizability in RNNs limits training efficiency. Transformer models, while enhancing global modeling capability through the self-attention mechanism, suffer from quadratic growth in computational complexity with sequence length. As the number of time steps increases, both computational and memory costs become major bottlenecks to training efficiency. In contrast, FCNN architectures adopt a fundamentally different non-recurrent feedforward structure, where gradients propagate only across network layers, thereby structurally avoiding the repeated gradient multiplication along time steps encountered in RNNs. To verify the rationality of the FCNN depth design and assess the model’s robustness to hyperparameters, this study conducted layer comparison experiments and hyperparameter sensitivity tests. These experiments were performed on the validation set to evaluate the impact of varying the number of FCNN layers (L = 2 to L = 6) on recommendation performance, as well as to explore the tuning ranges of learning rate (η), batch size, and task weight (α). The results are summarized in Table 8. Table 8Impact of FCNN layer number on model performance (mean ± standard deviation).Layers (L)AUC (%)NDCG@5 (%)Training Time (hours)Parameters (millions)292.1 ± 0.457.3 ± 0.33.2 ± 0.23.0393.5 ± 0.359.8 ± 0.23.8 ± 0.33.84 94.5 ± 0.3 61.0 ± 0.2 4.0 ± 0.3 4.5 593.9 ± 0.560.1 ± 0.44.5 ± 0.45.2692.7 ± 0.658.6 ± 0.55.1 ± 0.56.0 Table 8Impact of FCNN layer number on model performance (mean ± standard deviation).Layers (L)AUC (%)NDCG@5 (%)Training Time (hours)Parameters (millions)292.1 ± 0.457.3 ± 0.33.2 ± 0.23.0393.5 ± 0.359.8 ± 0.23.8 ± 0.33.84 94.5 ± 0.3 61.0 ± 0.2 4.0 ± 0.3 4.5 593.9 ± 0.560.1 ± 0.44.5 ± 0.45.2692.7 ± 0.658.6 ± 0.55.1 ± 0.56.0 Impact of FCNN layer number on model performance (mean ± standard deviation). As shown in Table 8, increasing the number of FCNN layers from 2 to 4 consistently improved model performance. Specifically, AUC and NDCG@5 increase by 2.4% and 3.7%, respectively, demonstrating that a deeper network enhances the model’s capacity to capture higher-order feature interactions. However, when the number of layers exceeded 4 (L = 5 or 6), performance started to decline, with the AUC decreasing by 1.8%. Meanwhile, both training time and the number of parameters increased significantly. This decline is attributed to the gradient vanishing problem common in deep networks, where high-level features lose information through multiple nonlinear transformations, and low-level features are not effectively updated during backpropagation. Consequently, L = 4 was chosen as the optimal depth, balancing performance, computational efficiency, and the ability to model complex user behavior features. Table 9Hyperparameter Sensitivity Analysis (L = 4 fixed).HyperparameterAUC (%)NDCG@5 (%)Training Stability (Loss Fluctuation)Learning Rate (η)1e-491.255.1High1e-3 94.5 61.0 Low 1e-289.753.8HighBatch Size6493.159.3Medium128 94.5 61.0 Low 25693.359.7MediumTask Weight (α)0.393.158.7Medium0.5 94.5 61.0 Low 0.793.058.5High Table 9Hyperparameter Sensitivity Analysis (L = 4 fixed).HyperparameterAUC (%)NDCG@5 (%)Training Stability (Loss Fluctuation)Learning Rate (η)1e-491.255.1High1e-3 94.5 61.0 Low 1e-289.753.8HighBatch Size6493.159.3Medium128 94.5 61.0 Low 25693.359.7MediumTask Weight (α)0.393.158.7Medium0.5 94.5 61.0 Low 0.793.058.5High Hyperparameter Sensitivity Analysis (L = 4 fixed). As shown in Table 9, the model achieved optimal performance when the learning rate was set to η = 1e-3. Both a smaller value (η = 1e-4) and a larger one (η = 1e-2) led to performance degradation due to unstable convergence or premature convergence to suboptimal solutions. In terms of batch size, 128 yielded the best trade-off between accuracy and training stability. A smaller batch size of 64 introduced excessive gradient noise, causing noticeable loss fluctuations, while a larger batch size of 256 reduced the parameter update frequency and failed to capture fine-grained behavioral signals. For the task weight α, which balances the primary recommendation objective with the interest distribution task, the value α = 0.5 produced the highest AUC and NDCG@5, indicating the strongest multi-task synergy. Deviating from this value (e.g., α = 0.3 or 0.7) disrupted the equilibrium between accuracy and diversity, which supports the theoretical rationale proposed in Sect. 3.3. Discussion Interpretation of the results The performance of the proposed MT-FCNN in dynamic interest modeling and diversity-aware recommendation can be theoretically explained from the joint perspectives of multi-task learning and behavior sequence analysis. According to the multi-task learning framework proposed by Li et al. (2022), shared representations across related tasks improve model generalization by introducing implicit regularization40. This principle is reflected in the collaborative optimization of MT-FCNN’s click prediction and interest distribution tasks, where the shared behavioral embedding layer encourages the extraction of generalizable interest patterns rather than overfitting to a single objective. In addition, the use of dynamic sequence embeddings in MT-FCNN addresses the issue of interest drift, as highlighted by Tang et al. (2025), who argue that user preferences evolve non-linearly over time and should be modeled through flexible feature interactions rather than static time-window statistics41. By applying high-order nonlinear transformations within fully connected layers, MT-FCNN transcends the limitations of RNNs, which tend to overemphasize local temporal dependencies, and instead captures global cross-behavioral semantic associations. This design aligns with the concept of long-term interest modeling discussed by Xu et al. (2023)42. Moreover, the superior performance of MT-FCNN in cold-start scenarios can be theoretically linked to the domain adaptation paradigm in transfer learning, wherein the integration of multimodal features—such as title semantics—facilitates knowledge transfer to users with sparse behavioral data, thereby reducing dependence on explicit interaction histories. Compared with recent related studies, Lyu et al. (2026) proposed a graph-coupled convolutional model that enhanced interpretability by assigning aggregation weights based on node similarity through an adaptive weighting mechanism43. However, this method still relied on explicitly constructed graph structures, leading to high update costs in dynamic news recommendation scenarios. Zeng et al. (2022) introduced a personalized recommendation algorithm based on GRU and attention mechanisms, which improved data completeness through a variable-length mini-batch allocation strategy44. Nevertheless, its modeling framework remains primarily recurrent in nature, limiting its ability to capture higher-order nonlinear interactions. Cao et al. (2025) conducted a theoretical analysis of dimensional requirements in non-contrastive self-supervised learning methods and demonstrated that performance was strongly influenced by the relationship between output dimensionality and latent class number. In particular, performance degraded significantly when the output dimension was substantially smaller than the number of latent categories45. This finding highlights the critical role of representation space design in recommendation performance. In contrast, MT-FCNN learns complex interest structures in a more flexible representation space through multi-layer fully connected transformations. Zhang et al. (2025) proposed a streaming data clustering method that integrated self-growing mapping with fast hierarchical merging. By enabling density-sensitive neurons to adapt to distribution shifts and rapidly merge cluster structures, the method effectively addressed dynamic clustering imbalance46. However, that method focuses on unsupervised clustering and differs from the end-to-end supervised optimization used in recommendation. These recent studies have improved recommendation performance from different perspectives, including graph-based modeling, sequential learning, self-supervised representation learning, and dynamic clustering. However, they still generally suffer from strong structural dependence, insufficient utilization of supervised signals, or limited capability in modeling higher-order nonlinear interactions. In contrast, MT-FCNN integrates behavior sequence modeling and multi-task optimization within a unified FCNN architecture. Without relying on explicit graph structures, it enables dynamic interest modeling and multi-objective learning, thereby offering more comprehensive advantages in scalability, modeling flexibility, and practical applicability. Implication The practical significance of this study lies in the dual optimization it offers to the news recommendation ecosystem. For users, dynamic interest modeling facilitates more accurate alignment with real-time information needs—for example, by automatically identifying a preference for short-form news during commuting hours and a tendency toward in-depth content consumption in the evening—thereby improving the coherence and overall satisfaction of the reading experience. For platforms, the diversity-aware recommendation enabled by the multi-task learning framework helps mitigate the dominance of head content, promotes the exposure of long-tail news items, and indirectly enhances the motivation of content creators, contributing to a more balanced and sustainable content ecosystem. Furthermore, the model’s relatively low computational complexity—compared to architectures based on self-attention mechanisms—offers practical feasibility for real-time deployment, addressing the news industry’s demand for high responsiveness and low latency. Limitations Despite the substantial advantages demonstrated by MT-FCNN, several limitations should be acknowledged. First, the model’s reliance on computational resources limits its applicability in edge environments or low-spec devices; for example, mobile applications must carefully balance computational overhead with real-time responsiveness and user experience. Second, while multimodal feature fusion improves generalization on sparse data, the semantic embedding of news titles depends on pretrained language models, which may introduce domain-specific adaptation biases. Third, the current framework does not explicitly model the temporal decay of news relevance (e.g., the rapid obsolescence of trending topics), which may reduce the model’s adaptability to rapidly evolving content. Lastly, the interpretability of MT-FCNN remains constrained to feature contribution analyses, lacking intuitive visualization tools to trace the evolution of user interests over time. Future work Future research may proceed along several promising directions. First, the adoption of lightweight model compression techniques, such as knowledge distillation or parameter pruning, could alleviate computational burdens and facilitate deployment on resource-constrained devices. Second, incorporating spatiotemporal contextual information—such as users’ geographical location and social interactions—into user profiling may enhance the system’s contextual sensitivity and personalization accuracy. Third, the integration of a temporal decay mechanism to dynamically modulate the relevance of news items based on recency could improve the system’s responsiveness to content volatility. Fourth, the development of interactive interpretability tools that enable users to adjust interest distribution parameters and observe real-time recommendation feedback may foster greater transparency and trust. Additionally, embedding an incremental learning framework to support online model updates would enhance the system’s adaptability to both the rapid evolution of news content and the continuous drift of user preferences. Conclusions This study proposed MT-FCNN, a personalized news recommendation model based on a fully connected neural network, to address challenges in dynamic interest modeling, recommendation performance, and cold-start scenarios. The model incorporates dynamic sequence embeddings and a multi-task learning mechanism to jointly optimize click intention prediction and interest distribution learning, thereby enhancing both recommendation accuracy and representation capability. Empirical validation on the MIND dataset confirms that MT-FCNN outperforms traditional User-CF, CBR-TFIDF, and existing deep learning baselines. This model achieves substantial improvements in AUC, NDCG, and CTR, and an 11.2% increase in recommendation accuracy for cold-start users, which underscores its robustness in data-sparse scenarios. In addition to accuracy, MT-FCNN advances recommendation diversity and user satisfaction by significantly improving coverage and novelty, thereby enriching content variety and fostering deeper user engagement. These outcomes highlight the model’s capability to integrate dynamic user behavior modeling with multi-task optimization, resulting in precise and personalized recommendations and heightened user satisfaction. The main contributions are threefold. First, a dynamic user behavior sequence modeling method and a unified representation learning framework based on a fully connected architecture are proposed. Second, a multi-task joint optimization framework is developed to mitigate the performance bias of single-objective learning. Third, the effectiveness of semantic feature integration is validated in data-sparse scenarios, confirming its auxiliary role in improving recommendation performance. Despite these achievements, several limitations remain. The model has relatively high computational complexity, requiring considerable resources for training and deployment, which may restrict its use on edge devices or under strict real-time constraints. It relies on pre-trained text embeddings, which may introduce domain bias and limit cross-domain generalization. Moreover, the model does not explicitly incorporate the temporal decay of news relevance, reducing its adaptability to rapidly evolving topics. Finally, the current interpretability analysis is mainly based on feature contribution scores and lacks an intuitive, dynamic representation of how user interests evolve over time. Future work will focus on model compression and lightweight design, such as parameter pruning and knowledge distillation, to reduce computational overhead and improve deployment efficiency. The integration of spatiotemporal contextual information and temporal decay mechanisms will be explored to enhance dynamic adaptation. Additionally, more interactive interpretability methods and incremental learning frameworks will be investigated to improve transparency and continual learning, facilitating large-scale real-world deployment. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1