RegionKG: A multimodal knowledge graph for comprehensive urban region representation and explainable urban computing

Urban region representation learning aims to learn intrinsic embeddings of urban areas from multi-source data for downstream applications. However, challenges remain in modeling of multimodal feature fusion and ensuring model interpretability. Therefore, this study proposes an urban region representation learning method based on multimodal knowledge graphs. First, the study designs RegionKGC method that integrates multimodal geographic data. Second, this study designs RegionKGF method for multi-source feature fusion using an adaptive weighting mechanism. Third, the study proposes RegionKGE, which integrates dynamic and static entity embedding modules, an entity enhancement embedding module supported by interwoven attention mechanisms, and a decision loss function combining hierarchy-aware and contrastive learning concepts to embed RegionKG into vector space. Finally, this study develops RegionKGX, an interpretability framework that leverages distance metrics between knowledge graph embedding vectors to extract evidence paths for downstream predictions. Experimental results demonstrate that RegionKGF and multimodal data can improve representation learning results, with RegionKGE showing improvements of at least 21% in HITS@1 compared to existing representation learning models such as MESN and SR-HAKE. Moreover, the learned embeddings achieve excellent performance on three downstream tasks: land use recognition, house price estimation, and urban village classification. The evidence paths from RegionKGX provide interpretable support for the model’s predictions and enable the evaluation of how spatial relationships impact different tasks. The proposed framework is of great significance for urban management decision support and the development of interpretable urban computing.