Exploring relevant snapshots and neighboring entities for temporal knowledge graph reasoning

Temporal Knowledge Graph Reasoning (TKGR) aims to predict missing entities at future timestamps based on evolving patterns in historical data. However, existing methods often over-rely on past events while neglecting the semantic evolution of future events, resulting in limited generalization to unseen facts. To address these challenges, we propose RSPED, a novel framework that enhances extrapolative reasoning by selecting relevant relational and temporal contexts. Specifically, we design: (1) a Relevant Relation Selector, which filters out irrelevant facts based on entity-relation dependencies and frequency-aware attention; and (2) a Potential Event Discovery module, which constructs auxiliary graphs to extract semantic dependencies among candidate entities. These two components enable the model to integrate both local and contextual signals across past and potential future snapshots. Extensive experiments on five public TKGR datasets demonstrate that RSPED consistently outperforms competitive baselines in terms of MRR (e.g., improving by 1.47%, 0.16%, 0.50%, 1.37%, and 3.88% on ICEWS14, ICEWS18, GDELT, YAGO, and WIKI, respectively), verifying its effectiveness and generalization in temporal reasoning tasks.