SHG: Self-Supervised Hypergraph Gate Networks for Enhanced Multi-Modal Representation

Graph Neural Networks (GNNs) have gained widespread popularity for encoding user-item interaction graphs and have consistently demonstrated strong performance. However, real-world user interaction data often exhibits noise, which can negatively impact the accuracy of GNN-based representations. In response to this challenge, we introduce a novel framework known as Self-supervised Hypergraph Gate Networks (SHG). This innovative approach employs a hypergraph-based encoder to learn hyperedge feature matrices. It utilizes a decoder with a multi-head attention mechanism to map hyperedge features back to entity features. Furthermore, we introduce a self-enhanced learning module that leverages global and local contexts to enhance the overall model performance. In comparison to baseline models, our model demonstrates exceptional performance across two publicly available multi-modal video recommendation datasets: TikTok and Movielens.