Learning Temporal Knowledge Graphs via Time-sensitive Graph Attention

Embedding-based graph representation learning methods have shown strong performance on knowledge graphs (KGs), but most are designed for static settings and struggle to model temporal dynamics. To address this gap, we present TimeGate, a novel time-sensitive graph attention network tailored for temporal KG (TKG) representation learning. Unlike traditional approaches that discretize graphs into time-based snapshots, TimeGate directly models timestamps as edge-level attributes and jointly encodes entities, relations, temporal signals, and neighbors into unified initial representations. A temporal-relational self-attention mechanism further refines these embeddings by adaptively weighting neighbors based on contextual relevance. A key innovation of TimeGate lies in its multi-model and multi-task generality: it supports both temporal KG completion and time-aware entity alignment (EA) within a unified, end-to-end architecture, and can flexibly serve as a plug-in encoder to boost diverse existing TKG models—without relying on recurrent networks or snapshot partitioning. Comprehensive experiments on four benchmark datasets (ICEWS, GDELT,Wikidata, and YAGO) demonstrate that TimeGate not only achieves state-of-the-art performance on time-aware EA tasks, but also consistently enhances the predictive accuracy of multiple backbone TKG completion models, highlighting TimeGate’s practical scalability, and broad applicability.