Abstract:
Spatial and temporal recommendation systems struggle to fully exploit the geographic context. However, information on administrative areas and relationships between points of interest remains largely underutilized. This thesis proposes to integrate an urban knowledge graph to fill this gap. Several prediction models are compared, such as: Each architecture is trained twice, without any graph information, and by injecting KG embeddings as static features or initial states. By combining symbolic reasoning (knowledge graph) and spatiotemporal deep learning, this study shows that an explicit geographic context significantly improves the accuracy of place recommendations.
