A Federated Social Recommendation Approach with Enhanced Hypergraph Neural Network

• DOI: 10.1145/3665931
• Publication Date: 2024-05-25T06:27:11Z
• Authors: Hongliang Sun; Zhiying Tu; Dianbo Sui; Bolin Zhang; Xiaofei Xu
• Keywords: Hypergraph; Social network (sociolinguistics)

In recent years, the development of online social network platforms has led to increased research efforts in recommendation systems. Unlike traditional systems, systems utilize both user-item interactions and user-user relations recommend relevant items, taking into account homophily influence. Graph neural (GNN)-based methods have been proposed model these item effectively. However, existing GNN-based rely on centralized training, which raises privacy concerns faces challenges data collection due regulations restrictions. Federated learning emerged as a privacy-preserving alternative. Combining federated with for can leverage their respective advantages, but it also introduces new challenges: (1) often lack capability process heterogeneous data, such relations; (2) sparsity distributed across different clients, capturing higher-order relationship information among users becomes challenging is overlooked by most To overcome challenges, we propose approach enhanced hypergraph (HGNN). We introduce HGNN learn user embeddings leveraging structure address heterogeneity data. Based carefully crafted triangular motifs, merge nodes construct hypergraphs local specific relations. Multiple channels are used encode categories high-order relations, an attention mechanism applied aggregate embedded from channels. Our experiments real-world datasets demonstrate effectiveness approach. Extensive experiment results three publicly available validate method.