E-MIGAN: Tackling Cold-Start Challenges in Recommender Systems

A recommender system based on Graph Neural Networks can effectively capture user-item interactions through the graph structure, leading to highly personalized and relevant recommendations. However, existing works adapting Graph Neural Networks (GNN) to recommendations struggle with the cold-start problem. Indeed, it is difficult to make accurate recommendations for new users or items with little or no interaction data. Building on previous work, we introduce an Enhanced Mutual Interaction Graph Attention Network (E-MIGAN) for this purpose. It is based on self-supervised representation learning on a large-scale bipartite graph. It is composed of three components: i) The attention network module that learns attention weights for each node and its neighbors, ii) The mutual interaction module computes a mutual interaction matrix for each node and its neighbors on each item, which encodes the pairwise interactions, and iii) A Content-Based Embedding model, which overcomes the cold start issue. The empirical study on real-world datasets proves that E-MIGAN achieves state-of-the-art performance, demonstrating its effectiveness in capturing complex interactions in graph-structured data.