Region-adaptive representation learning for geographic knowledge graph

Geographic knowledge graph (GeoKG) organizes the geographical entities and their relationships, which provides rich spatial semantic to serve various geographic artificial intelligence tasks by knowledge graph representation learning (KGRL). As a bridge between the knowledge graph and downstream tasks, KGRL results in a speedup of inference by embedding the entities and relationships of KG into a low-dimensional vector space. However, the existing KGRLs treated subgraphs in different regions uniformly, ignoring the spatial heterogeneity of GeoKG. It may result in significant performance differences across regions. To address this problem, this study proposes a region adaptive KGRL (RA-KGRL) method to mitigate the impact of spatial heterogeneity on performance. Specifically, RA-KGRL first splits the original GeoKG into subgraphs by introducing regional priors. Then, RA-KGRL learns local models for the subgraphs and performs local-to-global optimization to learn accurate global model. Extensive experiments on thirteen datasets indicate that RA-KGRL is comparable or even better than baselines on the traditional metrics, and outperforms all baselines on the new region adaptive metrics. This study provides a methodological reference for improving the performance of KGRLs on spatial heterogeneity scenarios.