Research On Collaborative Filtering Recommendation Algorithm Based On Graph Neural Networ

Nowadays,personalized recommendation systems are built based on massive data mining to recommend information and products of interest to users based on their historical information data,such as interest preferences and historical consumption behavior.Therefore,learning users’ personal preferences from their historical information has become a hot research task.However,as more and more products enter the market,user historical data has become increasingly complex,and recommendation tasks face great challenges.Traditional collaborative filtering-based recommendation methods can no longer meet the needs of current recommendation tasks.However,the development of graph neural network technology brings new opportunities.Nowadays,recommendation systems based on graph neural networks are widely used in many fields.It constructs user history information into a graph network structure,which can help it better understand and analyze users’ preferences.Based on this,this paper mainly studies collaborative filtering recommendation methods based on graph neural networks and proposes three different recommendation methods,namely:(1)Graph Initialization Learning Method Based on Graph Recommendation.Most existing collaborative filtering models based on graph convolutional networks suffer from two main issues.Firstly,the noise and data sparsity in the original graph significantly impairs the model’s performance.Secondly,for large user-item graphs,the explicit message passing in traditional GCN slows down the convergence speed during training,reducing the training efficiency of the model.To address these issues,a graph convolutional collaborative filtering model combining graph enhancement and sampling strategies is proposed,EL-GCCF.Firstly,the graph initialization module enhances the original graph by generating two types of graph structures that consider structural and feature information,effectively alleviating the noise problem.Secondly,the explicit message passing in traditional GCN is skipped using a multi-task constrained graph convolution.An auxiliary sampling strategy is used to effectively alleviate the problem of over-smoothing during training,improving the training efficiency of the model.(2)Neural Collaborative Filtering Model combining Attribute Awareness and Graph Matching.The attribute interactions between users and items can be divided into intrinsic interactions,which lie in learning the internal features of users or items,and crossinteractions,which learn the interaction features between users and items.However,previous recommendation models do not distinguish between these two interactions,which may lose much valid data when learning interaction information.A graph neural collaborative filtering model combining attribute awareness and graph matching was proposed to address the above problems,GNAGM.Firstly,users and items are represented as graph structures,and then the embedded representation of the graph is learned using neural networks.Second,a graph-matching algorithm is used to learn the similarities between users and items to make recommendations.By combining neural networks and graph matching algorithms,the method can better capture the complex relationships between users and items and has better recommendation performance.(3)Graph Convolutional Collaborative Filtering Model Combining Bi-directional Deep Context Awareness and Social Relationships.There are two under-studied problems with existing collaborative filtering models: First,for user interest,existing approaches only aggregate information about friends in candidate items,and these shallow contextual information aggregations make them plagued by limited information about friends.Secondly,for item attractiveness,such items may be more attractive if the past consumers of the items are friends of the target users or users with similar consumption habits.However,most current approaches ignore this deep context-aware information.A graph convolutional collaborative filtering model combining bi-directional deep context awareness and social relationships was proposed to address the above problems,BIDCA.Firstly,two graph structures are constructed by synergistic similarity relations.Secondly,the synergistic and social relationships on the synergistic similarity graph and the social network graph are modeled using a relationship-aware graph neural network to capture the higher-order relationships in the graph network structure,and based on this,the social information and the attractiveness of the items are captured using bidirectional deep context-awareness.