GeoKG-Enabled Similarity Computation with Defeasible Reasoning

This paper investigates the application of defeasible reasoning within geospatial knowledge graphs (GeoKGs) for geospatial similarity computation. Motivated by the need for accurate and interpretable similarity assessments in domains such as urban planning and location-based services, this study proposes a novel approach that combines the structured data representation of GeoKGs with the uncertainty-aware inference capabilities of defeasible logic. A GeoKG is constructed by integrating data from OSMnx, Wikipedia, and GeoNames. Defeasible rules are generated to capture contextual and functional similarities, and a reasoning engine infers similarity scores through priority-based conflict resolution. The proposed method is benchmarked against knowledge graph embedding (KGE) models and a large foundation model (Gemini gemini flash 2.0) using an expert-annotated dataset. While the KGE model achieved 72.3% accuracy and the LFM 68.1%, defeasible reasoning achieved 67.2%. Despite its lower accuracy, it offers superior interpretability by explicitly representing the rationale behind similarity assessments. This transparency is critical in decision-making scenarios where trust and justification are paramount. The study also highlights the impact of rule refinement and conflict resolution strategies on performance, suggesting potential for further improvement. By introducing defeasible reasoning into GeoKG-based similarity computation, this work provides a promising, explainable alternative to black-box models, paving the way for future hybrid approaches that balance accuracy and interpretability.