Entity Alignment for Temporal Knowledge Graphs

The School of EECS is hosting the following PhD Progress Review 3 Seminar:

Entity Alignment for Temporal Knowledge Graphs 

Speaker: Jiayun Li

Abstract: Temporal Entity Alignment (TEA) — the task of identifying equivalent entities across Temporal Knowledge Graphs (TKGs) — is essential for integrating knowledge from diverse sources. Despite recent progress, existing TEA models face three compounding challenges that have yet to be addressed in a unified manner. First, prior models treat temporal features uniformly during embedding learning, neglecting the varying importance of temporal frequency and the richness of neighbourhood context. Second, existing approaches fail to adaptively balance the orthogonal yet complementary contributions of structural and temporal features at the entity level, limiting their representational capacity under noisy and heterogeneous conditions. Third, and most critically, all existing TEA methods operate under a static snapshot assumption, making them ill-suited for real-world TKGs that evolve continuously over time. In this work, we present a series of progressively advancing frameworks to address these limitations. We first propose HTEA, a heterogeneity-aware embedding model that introduces frequency-based temporal adjacency matrices, a temporal richness attention mechanism, and an iterative refinement strategy for detecting and correcting temporally heterogeneous facts across TKGs. We then propose RCTEA, which further enhances TEA via richness-guided attention and adaptive entity-wise feature weighting, coupled with a dual-view neighbourhood consensus algorithm that jointly refines structural and temporal encoders for robust alignment. Finally, we propose ELITE, a lightweight continual TEA framework that extends alignment to the realistic evolving setting, where TKGs grow incrementally across snapshots. ELITE introduces dual-aspect complementary seed selection grounded in expected gradient length theory, localised inference restricted to evolving graph regions, and cross-snapshot score aggregation with confidence-margin-based conflict resolution. Extensive experiments on multiple TEA benchmarks — including the newly constructed YAGO-WIKI180K, BETA, YAGO-WIKI180K-E, and BETA-E datasets — demonstrate that our frameworks consistently achieve state-of-the-art alignment performance, with ELITE further reducing inference time by 4× and memory cost by 10× over full-graph baselines. 

Bio: Jiayun Li is a final-year Ph.D. student at the University of Queensland, Australia. Prior to his candidacy, he earned a Master of data science from UQ. Jiayun is currently working on the research topic of temporal entity alignment, advised by Prof.Xue Li and A/Prof.Wen Hua.