Region-adaptive representation learning for geographic knowledge graph

Area of Science:

• Geographic Information Science
• Artificial Intelligence
• Data Science

Background:

• Geographic knowledge graphs (GeoKG) are crucial for geographic artificial intelligence (GeoAI) tasks.
• Knowledge graph representation learning (KGRL) embeds GeoKG entities into vector spaces for faster inference.
• Existing KGRL methods overlook spatial heterogeneity, leading to performance disparities across regions.

Purpose of the Study:

• To propose a novel region adaptive KGRL (RA-KGRL) method.
• To mitigate the performance impact of spatial heterogeneity in GeoKG.
• To enhance the accuracy and applicability of KGRL in diverse geographic contexts.

Main Methods:

• The proposed RA-KGRL method partitions GeoKG into regional subgraphs using prior information.
• It learns localized models for each subgraph.
• A local-to-global optimization strategy is employed to refine a comprehensive global model.

Main Results:

• RA-KGRL demonstrates comparable or superior performance to baseline methods on traditional metrics.
• The method significantly outperforms existing approaches on newly introduced region-adaptive metrics.
• Experiments were conducted across thirteen diverse datasets.

Conclusions:

• RA-KGRL effectively addresses spatial heterogeneity challenges in GeoKG.
• The proposed method offers a valuable approach for improving KGRL performance in geographically diverse scenarios.
• This study provides a methodological reference for future research in spatial knowledge graph representation learning.