Personalized Recommendation System with Enhanced Graph Contrastive Learning based on Extra Information

• 1、School of Computer Science(National Pilot Software Engineering School), Beijing University of Posts and Telecommunications

Abstract：In recent years, Graph Neural Networks (GNNs) have been extensively researched and rapidly evolved in the field of recommendation systems. Traditional GNN-based recommendation systems construct bipartite graphs using user-item interaction data and employ GNNs to learn embeddings for users and items. Most research has focused on enhancing graph structures and convolution methods. However, relying solely on user-item interaction data leads to issues of data sparsity and overlooks several important aspects: 1) Users\' preferences are not only reflected in the user-item interaction graph but are also influenced by social relationships, such as user-user interactions and the groups users join. While multi-hop graph convolutions can capture some user-user relationship information, this information diminishes with each hop due to propagation and aggregation operations. Moreover, most models do not incorporate group information into personalized recommendations. 2) Users\' preferences can be inferred from the relationships between interacted items and other items, which is often overlooked. To better leverage user social information and item-item relationship information, this paper proposes a Graph Neural Network recommendation system - an Enhanced model Incorporating social and item relationship information along with Contrastive Learning for personalization (EIGCL). EIGCL utilizes user-item interaction data to mine user-user and item-item relationship information based on co-occurrence relationships. The model also leverages user-group information to derive group-item relationships. It constructs five types of graphs for each kind of relationship: User-Item graph (U-I), User-User graph (U-U), Item-Item graph (I-I), Group-User graph (G-U), and Group-Item graph (G-I). Convolutions are performed on these graphs to obtain embeddings for users, items, and groups. The model also adopts the successful contrastive learning approach used in recommendation systems to construct contrastive pairs from the embeddings derived from these five types of graphs. Extensive experiments on two real-world datasets demonstrate that EIGCL outperforms the latest state-of-the-art models.

Keywords： Personalized recommendation, graph convolution network, social network, item relationship, contrastive learning