Beyond Spatio-Temporal: Frequency-Gated Hypergraphs for Next-POI Recommendation
Abstract
Next Point-of-Interest (POI) recommendation aims to predict a user's subsequent location from historical trajectories. The task is challenging because user mobility is shaped by short-term intentions, long-term preferences, calendar effects, and structured spatial relations. Existing methods have improved sequential modeling and representation robustness, but they often treat temporal, spatial, and semantic contexts as loosely coupled signals and provide limited control over frequency-domain mobility patterns. To address these limitations, we propose ST-HFCR (Spatio-Temporal Hypergraph-enhanced Frequency-domain Contrastive Recommendation). ST-HFCR first constructs a multi-type hypergraph with time-slot, spatial-proximity, calendar-aware, and user-personalized hyperedges to model both macro-level contexts and individual preferences. It then introduces a gated frequency fusion module, GatedKAN, which learns separate gates for real and imaginary spectral components to balance amplitude- and phase-related information. Finally, ST-HFCR uses multi-level contrastive learning to align node- and graph-level representations under semantic-aware perturbations. Experiments on three real-world datasets show that ST-HFCR achieves consistent gains over competitive baselines in both HR and NDCG, suggesting that structured spatio-temporal hypergraphs and frequency-aware gated fusion provide complementary signals for next-POI recommendation. The code is available at https://github.com/MyCode12-maker/STHGCN/.