Deep Learning Meets Graph Theory: A Novel Approach to Generating Embeddings for Recommender Systems

Abstract

Graph-based models have emerged as a promising approach for capturing intricate relationships in various real-world applications [1], including recommendation systems. A novel approach called P-GNN (Partitioned-Graph Neural Network) has been proposed in this work, which utilizes the bipartite graph structure commonly used in recommendation systems. The approach generates separate embeddings for users and items, allowing highly discriminative recommendations. A significant advantage of this approach is its ability to address the oversmoothing and cold-start problem by computing separate subgraphs for users and items. This feature enables the generation of embeddings even for new users or items. The effectiveness of the proposed approach has been evaluated on multiple benchmark datasets and compared against state-of-the-art methods. The results demonstrate that Partitioned-GNN outperforms existing approaches, highlighting the potential of graph-based methods for recommendation systems.