Disentangled Hypergraph Network with Implicit Structure Learning for Mobility Social Relationship Inference

Abstract

Inferring social relationships from users' mobile data holds significant value for personalized recommendations. Most methods model user interactions based on co-occurrence records, achieving impressive success in capturing social signals. However, despite these advancements, current techniques still face three primary challenges. First, compressing higher-order co-occurrences into a single representation space easily leads to semantic entanglement. Second, relying on deterministic topology overlooks unobserved homophily. Third, the intrinsic characteristics of the users themselves are frequently ignored. To address these challenges, we propose a Disentangled Hypergraph Network with Implicit Structure Learning for mobility social relationship inference, named DHISL. Specifically, it first captures individual spatiotemporal characteristics through feature-guided representation initialization, and then adopts a dual-branch framework: the explicit branch models social and spatiotemporal preference signals, while the implicit branch adaptively learns a sparse channel-wise relational structure via differentiable structure learning. Furthermore, we introduce dual constraints in a multi-channel disentangled space to achieve intra-channel consistency and inter-channel exclusivity. Experiments on four real-world datasets demonstrate that DHISL outperforms state-of-the-art models, achieving average improvements of 6.34% in ROCAUC and 6.39% in PRAUC. The source code is available at https://github.com/Tilamisu-zz/DHISL.