Research And Implementation Of Recommendation System Based On Multi Graph Neural Networks

Recommendation system can help users make simpler decisions,improve user experience and increase viscosity of users to the platform.Collaborative Filtering(CF)is one of the most important recommendation techniques,but it faces two major challenges: data sparsity and cold start.While fusing side information into recommendation models can alleviate such problems,most models have difficulty capturing user-item historical interaction information.In addition,existing recommendation models cannot flexibly integrate multiple types of structured side information and cannot adequately capture heterogeneous information between entities.In this thesis,we propose a recommendation system framework based on multi-graph neural networks with the powerful propagation and representation capabilities of graph neural networks(GNNs),and improve some modules and techniques.The main work accomplished in this thesis is as follows:(1)A novel recommendation framework for multi-source side information fusion,GSICF,is proposed.GSICF consists of several parts,including graph data construction,model construction,model evaluation,discussion and analysis.Experiments show that GSICF can effectively and flexibly integrate user social networks,user-item interaction data and item structured attributes,and has good applicability,scalability and interpretability.The various types of side information can interact with each other and enhance the performance of the recommendation system.In addition,the framework can fully capture the structured side information of users and items,effectively extract the higher-order interactions of users and items,and make up for the shortcomings of existing methods in representing user interrelationships.(2)A Trans D-based embedding module is added to GSICF.In order for the recommendation model to learn better user and item embedding representations,we add a Trans D-based embedding module to the model.The module not only serves as a pre-training,but also can build multi-task learning models together with GNN.At the same time,the embedding module can distinguish different connection relationships between entities and capture heterogeneous information in heterogeneous graphs.The embedding module and the GNN propagation module complement each other and jointly contribute to the performance of the recommender system.Other works in this thesis include: two new GNN aggregators designed for GSICF that can propagate richer neighborhood information;designing a graph attention mechanism for GSICF to distinguish the importance of different neighbor nodes for specific relationships;performing linear propagation in the feature propagation process to make the model simpler;using a new prediction layer,i.e.,recording the embedding representation of nodes in each layer of GNN,as a way to alleviate the over-smoothing problem of GNN-based recommendation model and facilitate the capture of user potential interest at a distance.