Geometric Feature Enhanced Knowledge Graph Embedding and Spatial Reasoning

Lei Hu,Wenwen Li,Yunqiang Zhu

2024-10-24

Abstract:Geospatial Knowledge Graphs (GeoKGs) model geoentities (e.g., places and natural features) and spatial relationships in an interconnected manner, providing strong knowledge support for geographic applications, including data retrieval, question-answering, and spatial reasoning. However, existing methods for mining and reasoning from GeoKGs, such as popular knowledge graph embedding (KGE) techniques, lack geographic awareness. This study aims to enhance general-purpose KGE by developing new strategies and integrating geometric features of spatial relations, including topology, direction, and distance, to infuse the embedding process with geographic intuition. The new model is tested on downstream link prediction tasks, and the results show that the inclusion of geometric features, particularly topology and direction, improves prediction accuracy for both geoentities and spatial relations. Our research offers new perspectives for integrating spatial concepts and principles into the GeoKG mining process, providing customized GeoAI solutions for geospatial challenges.

Artificial Intelligence

What problem does this paper attempt to address?

The problem that this paper attempts to solve is that the existing knowledge graph embedding methods lack geographical awareness when dealing with geographical entities and their relationships, resulting in an inability to accurately reflect geospatial relationships in link prediction tasks. Specifically, traditional knowledge graph embedding techniques (such as KGE) mainly rely on the semantic structure of triples and ignore the geometric features (such as topology, direction, and distance) between geographical entities. This may make the selection of candidate entities in downstream tasks not in line with geographical intuition, thereby affecting prediction accuracy.

To meet this challenge, the author proposes a new method to enhance the general KGE model by injecting geographical geometric features into the knowledge graph embedding process. Specifically, geometric features between geographical entities, such as topology, direction, and distance, are introduced during the embedding process to reduce the distance between spatial relationship terms and their corresponding geometric features, so that spatial relationship terms that describe the same geometric configuration but are expressed in different languages are closer in the embedding space. This method aims to improve the effectiveness of KGE in geographical knowledge graphs, especially in link prediction tasks.

Through experimental verification, this method performs well in link prediction tasks for both geographical entities and spatial relationship terms. In particular, when topological and directional features are included, the prediction accuracy of the model is significantly improved. This shows that combining geometric features can effectively improve the spatial reasoning ability of geographical knowledge graph embedding models.