Proceedings of the Annual Meeting of the Cognitive Science Society

Predicting human mobility, specifically next Point-of-Interest (POI) suggestions, requires modeling how individuals integrate spatial, temporal and category cues during decision-making. However, existing computational models often conflate these heterogeneous signals or struggle with data sparsity. We propose the Disentangled Spatio-Temporal and Category Graph Network (DST-GN), a framework that reconstructs cognitive maps through representation learning. DST-GN models spatial, temporal and category contexts as independent graph views, employing Graph Attention Networks (GAT) to learn view-specific embeddings. To ensure coherence across these disentangled pathways, we introduce a cross-view contrastive learning mechanism that aligns signals into a unified semantic space. Furthermore, a global Collective Flow Enhancer is integrated to mitigate data sparsity by leveraging population-level heuristics. Experimental results on three real-world datasets show that our model outperforms state-of-the-art techniques, achieving average improvements of 22.84% in HR@10 and 18.55% in NDCG@10.