<p>Landslides are among the most widespread and destructive geological hazards in China, making landslide spatial prediction essential for refined disaster risk management. Although machine learning methods have achieved notable performance in landslide susceptibility assessment, they often inadequately represent mechanism knowledge, tend to focus on local-scale landslides and their surrounding environments, and strongly depend on training sample quality. Knowledge graphs provide a promising framework for systematically modeling landslide formation mechanisms and capturing latent relations among related entities. In this study, a landslide spatial prediction method, termed GeoSem-GraphSAGE, was proposed by integrating knowledge graphs and representation learning. A landslide knowledge graph was first constructed to explicitly model conditioning factors and their interaction relations, and semantic embeddings of graph nodes were learned using the ComplEx model. A directed weighted graph was established to capture geographical similarity beyond distance-based constraints, and the Louvain algorithm was applied for geographical partitioning to optimize non-landslide sample selection. Ultimately, the Graph sample and aggregate model was employed to jointly learn semantic information and graph structural features, enabling landslide susceptibility prediction. Experiments conducted in Anxi County demonstrated that the GeoSem-GraphSAGE outperformed representative non-graph-based models, achieving an accuracy of 90.22%, precision of 91.10%, recall of 89.18%, F1-score of 90.11%, and an AUC of 95.77%. Ablation results further indicated that explicit modeling of mechanism knowledge improved the delineation of susceptibility areas, while the geographical similarity–based sampling strategy enhanced sample quality, leading to an overall performance improvement of approximately 2%–4%. This approach provides effective support for regional landslide risk mitigation.</p>