STSE: Spatio-temporal state embedding for knowledge graph completion
Xiaoyu Ji, Cao Yibing, Zhang Jiangshui, Xinke Zhao · Knowledge-Based Systems · 2025
The explicit integration of temporal and geospatial features into knowledge graphs enables more precise characterization of knowledge dynamics across temporal and geographic dimensions, while simultaneously amplifying the complexity of inferring missing facts in spatio-temporal knowledge graphs (STKGs). To address these dual challenges in Spatio-Temporal Knowledge Graph Completion (STKGC), we present STSE (Spatio-Temporal State Embedder), an innovative embedding framework that systematically coordinates relational semantics with spatial-temporal continuum modeling. Our technical contributions manifest through three key innovations: (1) A novel descriptive form that enhances practicality by distinguishing head/tail entity locations in tuples; (2) A geometry-aware embedding space that dynamically fuses spatial grids and temporal slices through spatio-temporal states, enabling robust reasoning on heterogeneous graphs; (3) An enhanced encoder that captures complex spatio-temporal contextual relationships between entities using modified Transformer and ResNet architectures. Experimental validation across three benchmark datasets (YAGO11k-ST, Wikidata12k-ST, ICEWS05-ST) demonstrates STSE's superiority, achieving 3.8 % Mean Reciprocal Rank ( MRR ) improvement in time prediction and 7.1 % Hits@10 enhancement in spatial location reasoning compared to state-of-the-art baselines. This methodological breakthrough establishes three implementable principles for STKGC: (1) geometric consistency constraints during fusion operations, (2) spatio-temporal difference capturing across heterogeneous relationships, (3) cross-domain applications ranging from epidemiological spread prediction to urban mobility pattern mining. © 2025 Elsevier Science. All rights reserved