Entity class-restricted twin attention-based aggregation for geographic knowledge graph inference

Geographic knowledge graphs (GeoKGs) describe and simulate geographic phenomena by storing geographic entities with their semantic relations, providing a foundation for many real-world tasks such as urban function identification. However, GeoKGs are often incomplete due to manual construction, leading to inaccurate and/or suboptimal decision-making. This situation requires GeoKG inference (GeoKGI), which involves learning new knowledge by projecting entities and relations into embedding spaces to infer missing facts and complete GeoKGs. Nevertheless, current inference methods mainly focus on modeling universal knowledge graphs (KGs), which cannot be directly transferred into GeoKGs as, unlike KGs, their entities contain specific class information. Inference methods on universal KGs lack a clear manner of modeling such entity class information. Moreover, they often ignore the representation learning of relations, resulting in inadequate modeling of GeoKGs. To address this issue, we propose a novel entity class-restricted twin attention-based aggregation (ECTAA) framework for GeoKGI. Briefly, we first restricted entities by encapsulating the class information within their vector representations to aid the model in better understanding and distinguishing between different entities and their attributes. We then designed a twin aggregation module with attention mechanisms to adaptively incorporate context information into both entity and relation representations. This module simultaneously considers the impact of context on entities and relations, further promoting the capture of GeoKG structural information. Experimental results on five GeoKG datasets validated the effectiveness of ECTAA compared to several state-of-the-art universal KG inference methods in predicting geographic facts.