Research on toponym Knowledge Graph Entity Alignment Based on Spatio-temporal Heterogeneous-aware Attention Mechanism

Abstract

Toponym knowledge graphs (PNKGs) are pivotal in applications such as smart cities, yet the heterogeneity of multi-source data significantly hampers the integration and utility of geographic knowledge. Additionally, few existing approaches account for temporal dynamics in PNKGs. To address these limitations, this study proposes STGEA, an entity alignment method that integrates spatial and temporal relationships within toponym knowledge graphs. Initially, a pre-trained model generates semantic-aware embeddings to capture cross-graph entity similarities, then this study designs a heterogeneous graph transformer equipped with temporal relationship awareness, this transformer employs an attention mechanism to integrate spatial, temporal, and attribute information across toponyms. Experimental results show that STGEA outperforms state-of-the-art methods across key metrics: Hits@1 (94.8%, +3.6%), Hits@5 (98.1%, +2.7%), and MRR (0.96, +0.3). Ablation studies further demonstrate the critical role of spatial and temporal relationships in entity alignment tasks, with performance degradation observed when either relationship is removed (Hits@1 drops to 90.2% and 90.8%, respectively).