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 (Transferable Inductive Reasoning), 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.
Figures & tables
Figure 1: Illustration of Transductive vs. Inductive Reasoning on Emerging Entities.
Figure 2: (a) Entity emergence over time. Across four TKGs, new entities continuously emerge; about ≈25% of entities are unseen during training. (b) Performance comparison. Under Vanilla vs. Emerging settings, strong baselines consistently drop on emerging entity triples. (c) Representation collapse. On ICEWS14, t - SNE of LogCL shows representation collapsing after training, while known entities drift to a separate manifold. (d) Transferable patterns. Semantically similar entities share relation-conditioned patterns, enabling transfer to emerging entities.
Figure 3: The overall architecture of proposed TransFIR .
ICEWS14
ICEWS18
ICEWS05-15
GDELT
Method
MRR
Hits@3
Hits@10
MRR
Hits@3
Hits@10
MRR
Hits@3
Hits@10
MRR
Hits@3
Hits@10
CyGNet(2021)
0.0111
0.0098
0.0202
0.0031
0.0020
0.0047
0.0031
0.0020
0.0048
0.0067
0.0031
0.0147
REGCN(2021)
0.1175
0.1263
0.2232
0.0947
0.1004
0.1797
0.0887
0.0961
0.1803
0.0222
0.0209
0.0393
HiSMatch(2022)
0.0284
0.0285
0.0418
0.0055
0.0058
0.0076
0.0242
0.0238
0.0443
0.0159
0.0141
0.0270
MGESL(2024)
0.0309
0.0361
0.0603
0.0747
0.0809
0.1031
0.1069
0.1166
0.1563
0.0516
0.0471
0.0815
LogCL(2024)
0.1354
0.1770
0.2273
0.0903
0.1064
0.1548
0.1917
0.2452
0.2855
0.0473
0.0479
0.0973
Table 1: Performance comparison of inductive reasoning on emerging entities on four benchmarks. In each column, best results are highlighted in bold and second-best are underlined . For generative model GenTKG, MRR is unavailable due to it’s reliance on multiple generations for each query.
Figure 4: (a) t-SNE visualization showing the improved separation of clusters in TransFIR compared to LogCL, with Collapse Ratio improvement from 0.0055 to 0.8677. (b) Case analysis of three representative clusters and how TransFIR transfers reasoning patterns to emerging entities.
Figure 5: Ablation study results on four benchmarks, showing the performance impact of removing different components. Reported results are for Hits@10, with additional metrics in Appendix F.2 .
Figure 6: Experiment results on ICEWS14 under the Unknown and Emerging settings.
PLM
ICEWS14
ICEWS18
GDELT
Baseline
0.2273
0.1797
0.1131
T5
0.3057
0.2061
0.2082
RoBERTa
0.2934
0.1939
0.2289
Qwen3
0.2567
0.2009
0.2030
BERT
0.3246
0.2324
0.2278
Table 2: Different Textual Encoder experiment result on TransFIR .
Figure 7: GPU memory usage and empirical running time on ICEWS14. TransFIR achieves significantly lower peak GPU memory usage while maintaining competitive training speed.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Entities
Relation
Time Snapshots
Total Triples
Emerging Entities
ICEWS14
7128
230
365
90730
1301
ICEWS18
23033
256
304
468558
3434
ICEWS05-15
10488
251
4017
461329
1954
GDELT
7691
240
2976
2278405
1020
Appendix
Table 3: Statistics of all datasets, including ICEWS14, ICEWS18, ICEWS05-15 and GDELT.
Figure 8: t-SNE visualizations comparing LogCL and TransFIR on multiple datasets. The top row shows LogCL embeddings, with a clear representation collapse for emerging entities (red). The bottom row shows TransFIR , where emerging entities are well-separated into latent semantic clusters, significantly improving the Collapse Ratio across all datasets.
Figure 9: Ablation results on four benchmarks under MRR (top row) and Hits@3 (bottom row). The ranking of variants mirrors the main-paper Hits@10.
Figure 10: Results on ICEWS18, ICEWS05-15, and GDELT under the Unknown (blue; observable G<t ) vs. Emerging (green; zero history) settings. All methods improve with pre-query history, and TransFIR remains best on both MRR and Hits@10 across datasets.
Figure 11: Experiment results on ICEWS14 and ICEWS05-15 under different time splits.
Figure 12: Hyperparameter study on four benchmarks, exploring effects of codebook size K , Interaction Chain length k , hidden dimension d , and the number of layers l . Brown and green represent MRR and HITS@10, respectively.
Temporal knowledge graph (TKG) reasoning aims to infer future facts at unseen timestamps from temporally evolving entities and relations. Despite recent progress, existing approaches still suffer from inherent limitations due to their inductive biases, as they predominantly rely on time-invariant or weakly time-dependent structures and overlook the evolutionary dynamics. To overcome this limitation, we propose a novel collaborative learning framework for TKGR (dubbed CID-TKG) that integrates evolutionary dynamics and historical invariance semantics as an effective inductive bias for reasoning. Specifically, CID-TKG constructs a historical invariance graph to capture long-term structural regularities and an evolutionary dynamics graph to model short-term temporal transitions. Dedicated encoders are then employed to learn representations from each structure. To alleviate semantic discrepancies across the two structures, we decompose relations into view-specific representations and align view-specific query representations via a contrastive objective, which promotes cross-view consistency while suppressing view-specific noise. Extensive experiments verify that our CID-TKG achieves state-of-the-art performance under extrapolation settings.
Shuai-Long Lei, Xiaobin Zhu, Jiarui Liang +3
University of Science and Technology Beijing, Haidian, Beijing, China
Temporal knowledge graphs (TKGs) represent time-stamped relational facts and support a wide range of reasoning tasks over evolving events. However, existing methods produce entity representations that are static at the entity level, in that each representation is a function of learned parameters only and retains no trace of the interactions in which the entity has participated. In this paper, we depart from this static view and propose that each entity be modeled as an adaptive process whose representation is refined every time the entity participates in a fact. To this end, we propose AdaTKG, which maintains a per-entity memory that is updated with every observed interaction, with the memory accumulating online and predictions improving as more interactions arrive. Specifically, we instantiate the memory update as a learnable exponential moving average governed by a single shared scalar instead of using learnable parameters for each entity, enabling AdaTKG to handle entities unseen during training. Extensive experiments confirm consistent gains over TKG baselines, demonstrating the effectiveness of adaptive memory. Code is available at: https://github.com/seunghan96/AdaTKG
Temporal Knowledge Graph (TKG) reasoning under the extrapolation setting focuses on forecasting future time-stamped events (facts) from historical data in a temporal knowledge graph. Existing approaches, reinforcement learning (RL)-based multi-hop reasoning methods are prominent for TKG reasoning because they produce human-interpretable predictions via explicit multi-hop path tracing. However, during RL training, rewards are typically sparse, and exploration is highly inefficient due to the vast, time-evolving action space. These issues hinder efficient training and often limit overall performance. To address these challenges, we propose RAPTOR (Reachability-Aware Pretraining for Efficient Target-Oriented Path Exploration), a self-supervised pretraining method that injects a reachability-aware inductive bias to the agent. By learning to estimate the reachability of candidate actions to the target entity, RAPTOR reduces exploration over unpromising paths and provides a strong initialization for downstream RL fine-tuning. Experimental results on the ICEWS14, ICEWS05-15, and ICEWS18 datasets demonstrate that RAPTOR pretraining markedly improves the training efficiency and consistently outperforms conventional baselines, establishing it as an effective approach for enhancing RL-based multi-hop reasoning methods for TKG reasoning.
Chien-Liang Liu, Tsao-Lun Chen
Artificial Intelligence Research Center, Chang Gung University · 2Graduate Institute of Oral Biology, National Taiwan University · 3Electrical Engineering, National Taiwan University of Science and Technology