GC-GAT: An integrated approach combining grid calculation and graph attention network for geographic knowledge graph reasoning

Abstract. Geographic knowledge graph (KG) reasoning enables the inference of missing entities and relations, a fundamental step toward constructing comprehensive geospatial knowledge systems. However, prevailing approaches often struggle with the accurate modeling of complex spatio-temporal dependencies, particularly in the context of large-scale data, and frequently lack interpretability. To address this issue, we present GC-GAT, a hybrid framework that integrates grid-based spatial encoding with a graph attention network to enhance geographic KG reasoning. We benchmark GC-GAT against other 9 models, including TransE, DistMult, ConvE, ComplEx, R-GCN, TANGO-DistMult, RE-NET, CyGNet, and RE-GCN. Our evaluation is conducted on the ICEWS14s dataset. Our framework significantly enhances the performance of the baseline on the spatio-temporal entity reasoning, achieving MRR of 46.11. Our results show that GC-GAT achieves superior accuracy and training efficiency, benefitting from the multiscale spatial aggregation properties of GeoSOT. These findings demonstrate the robustness and scalability of GC-GAT for reasoning over geographic knowledge graphs.