Inductive Reasoning for Temporal Knowledge Graphs with Emerging Entities Ze Zhao, Yuhui He, Lyuwen Wu, Gu Tang, Bin Lu, Xiaoying Gan, Luoyi Fu, Xinbing Wang, Chenghu Zhou Abstract Reasoning on Temporal Knowledge Graphs (TKGs) is essential for predicting future events and time-aware facts. While existing methods are effective at capturing relational dynamics, their performance is limited by a closed-world assumption, which fails to account for emerging entities not present in the training. Notably, these entities continuously join the network without historical interactions. Empirical study reveals that emerging entities are widespread in TKGs, comprising roughly 25% of all entities. The absence of historical interactions of these entities leads to significant performance degradation in reasoning tasks. Whereas, we observe that entities with semantic similarities often exhibit comparable interaction histories, suggesting the presence of transferable temporal patterns. Inspired by this insight, we propose TransFIR ( Transf erable I nductive R easoning), a novel framework that leverages historical interaction sequences from semantically similar known entities to support inductive reasoning. Specifically, we propose a codebook-based classifier that categorizes emerging entities into latent semantic clusters, allowing them to adopt reasoning patterns from similar entities. Experimental results demonstrate that TransFIR outperforms all baselines in reasoning on emerging entities, achieving an average improvement of 28.6% in Mean Reciprocal Rank (MRR) across multiple datasets. The implementations are available at https://github.com/zhaodazhuang2333/TransFIR . 1 INTRODUCTION Reasoning on Temporal Knowledge Graphs(TKGs) facilitates the prediction of future events and time-aware facts, significantly enhancing the utility and applicability of temporal knowledge graphs. By explicitly modeling relation dynamics as the graph evolves, TKG reasoning captures temporal dependencies and interaction patterns, thereby supporting event forecasting and time-aware inference ( 22 ; 46 ) . These capabilities form the foundation for applications such as temporal question answering, clinical risk analysis, and recommendation systems ( 44 ; 30 ; 13 ) . However, existing reasoning methods focus on modeling relation dynamics while neglecting the emergence of new entities. In real-world graphs, both entities and relations evolve continuously. Emerging entities often join the network without historical interactions . This phenomenon is observed in various contexts, from social platforms adding new users ( 38 ) to molecular networks coming new compounds ( 12 ) . Although current methods effectively capture relation dynamics and achieve strong forecast performance ( 21 ; 43 ) , they typically assume a closed entity set. Due to the absence of historical interactions, these models lack adequate supervision and representation for emerging entities, which significantly limits their reasoning capability. For instance, as shown in Fig. 1 , when Barack Obama first assumes office, predicting his first state visit becomes challenging due to the absence of historical interactions. To clarify the challenges and opportunities for reasoning on emerging entities, we conduct an empirical study from three progressively deeper perspectives (see Sec. 3). From the Data perspective, we observe that emerging entities are widespread in TKGs, with nearly 25% of entities not appearing in the training set. Meanwhile, existing models show a significant performance drop on related events. From the Representation perspective, we attribute this degradation to representation collapse , caused by the lack of supervision signal from historical interactions . Finally, from the Feasibility perspective, we explore invariant patterns to transfer to emerging entities and find that entities of semantically similar type often exhibit comparable interaction histories. As a inspiring method, inductive learning provides a promising approach for reasoning on new entities in knowledge graphs. Unlike transductive methods, which rely on entity-specific embeddings, inductive approaches learn transferable patterns from subgraphs ( 5 ) . For instance, InGram ( 18 ) constructs relation-affinity graphs to capture neighbor interactions, while ULTRA ( 10 ) generalizes to unseen entities through relative interaction representation. However, these methods are primarily designed for static KGs, where new entities already have known interactions. In contrast, emerging entities in TKGs often arrive without any interactions. This lack of supervision signals can lead to representation collapse, raising a significant challenge: how can we prevent representation collapse in the absence of historical interactions. Figure 1: Illustration of Transductive vs. Inductive Reasoning on Emerging Entities. To address this challenge, we propose TransFIR ( Transf erable I nductive R easoning), an inductive reasoning framework designed to handle emerging entities in TKGs. Inspired by our empirical observation that semantically similar entities exhibit transferable patterns, we propose Interaction Chain to model such structures. TransFIR extracts these patterns from Interaction Chains and employs a codebook-based classifier to map entities into latent semantic clusters, thereby transferring patterns to emerging entities. Specifically, TransFIR follows a Classification–Representation–Generalization pipeline: (i) Classification maps entities to latent semantic clusters via an interaction-aware codebook; (ii) Representation encodes entity’s Interaction Chain to capture reasoning patterns; (iii) Generalization propagates learned patterns within each cluster, enabling emerging entities to obtain informative embeddings. Together, these steps help prevent representation collapse and improve forecasting performance for emerging entities. In summary, the main contributions of our work are as follows: • Novel Framework. We propose TransFIR , an inductive framework designed to transfer reasoning patterns from semantically similar entities to enable reasoning on emerging ones. • Codebook-based Classifier for Transfer. We propose an interaction-aware VQ codebook that maps entities into latent semantic clusters. This facilitates reasoning pattern transfer while preventing representation collapse. • Problem & Evidence. We formally define the task of inductive reasoning on emerging entities without historical interactions. Additionally, an empirical study demonstrates that such entities are widespread in TKGs and existing methods suffer significant performance degradation. • State-of-the-art Results. TransFIR outperforms strong baselines across multiple benchmarks, with an average improvement of 28.6\% in MRR on four datasets. 2 Preliminaries and Problem Formalization Reasoning on Temporal Knowledge Graphs. A temporal knowledge graph (TKG) is structured as a sequence of timestamped snapshots \mathcal{G}=\{\mathcal{G}_{t}\}_{t\in\mathcal{T}} , where each snapshot \mathcal{G}_{t}=(\mathcal{E}_{1:t},\mathcal{R},\mathcal{F}_{t}) . Here \mathcal{E}_{1:t} is the set of entities observed up to time t , \mathcal{R} is the relation set, and \mathcal{F}_{t}\subseteq\mathcal{E}_{1:t}\times\mathcal{R}\times\mathcal{E}_{1:t}\times\{t\} represents the set of timestamped facts. Given a query (e_{s},r,?,t_{q}) with t_{q} in the future, temporal KG reasoning aims to predict the missing entity based on the historical context \mathcal{H}_{t_{q}}=\bigcup_{i0 as weighting coefficients. This optimization refines the codewords into semantically coherent clusters and stabilizes the assignment of entities to prototypes. Unlike static clustering methods, our approach jointly learns the prototypes with the task objective, making it suitable for fixed entity embeddings while enabling adaptive category representations for effective classification. 4.2 Interaction Chain Encoding (Representation) To capture transferable interaction sequence patterns for emerging entities, we introduce an Interaction Chain (IC) around each query entity. Unlike unordered temporal neighborhoods, ICs preserve the sequential structure of interactions, thereby reflecting entity-invariant temporal dynamics—such as periodic behaviors or events that follow a specific order. Definition. Given a temporal query q=(e_{q},r_{q},?,t_{q}) and a window size T , the IC of e_{q} collects its past interactions in chronological order: C_{q}=\big\{(e_{q},r,o,t_{i})\ \text{or}\ (s,r,e_{q},t_{i})\ \big|\ t_{q}-T\leq t_{i}