ABSTRACT Geographic knowledge graph representation learning (GeoKGRL) seeks to embed entities and relationships within a geographic knowledge graph (GeoKG) into a vector space, enabling efficient computation, retrieval, and inference on large‐scale datasets. However, existing GeoKGRL methods often overlook the hierarchical nature of geographic entities such as transboundary watersheds and nested administrative regions, leading to inadequate modeling of cross‐hierarchy distributional divergence. Consequently, the learned representations may not fully capture the inherent geographic semantics of geographic entities. To address this challenge, we propose a Hierarchical Adaptive Meta‐Learning (HAML) approach for enhancing GeoKGRL. Specifically, HAML introduces a region‐guided subgraph sampling strategy and a subgraph learning module to effectively capture entity and relationship characteristics across different geographic scales. Additionally, a hierarchical optimization module is designed to refine geographic knowledge representations through a “bottom‐up update, top‐down feedback” mechanism. By dynamically extracting local structural patterns at varying geographic scales, HAML enhances representation robustness, while a global meta‐optimizer facilitates cross‐hierarchy learning, mitigating distributional divergence between hierarchical embeddings. Extensive experiments conducted on OpenStreetMap‐based datasets demonstrate that HAML significantly outperforms baseline methods. These results highlight the effectiveness of HAML in explicit modeling of GeoKG structures, offering valuable insights for both GeoKG research and broader geographic applications.
Yao et al. (Fri,) studied this question.