Recommendation Algorithm Based On Graph Neural Network

Traditional recommendation algorithms have insufficient ability to model complex useritem interaction data.Most models based on single user behavior do not consider the information contained in other auxiliary behaviors,and have strong dependence on historical interaction data,resulting in serious cold start problems..The graph network is closer to the real network structure,and the semantic relationship mining of the rich node and edge information can improve the effect of the recommendation algorithm.At present,the model based on graph neural network has the following problems: With the development of the Internet,many new users join and many novel items appear,which makes the cold-start problem of the recommendation algorithm serious,and the cold-start neighbors are not considered when aggregating the message of the target node.Node processing affects the embedding quality of target nodes;the dynamics and complex dependencies of user-item interaction behaviors are ignored;social recommendations integrated into user-side knowledge graphs face the problem of social inconsistency.To this end,this paper proposes a pre-trained multi-behavior recommendation algorithm Pre-MBRec(Pre-trained Multi-behavior Recommendation based on GNN)based on graph neural network.The algorithm is divided into three stages.First,the graph neural network is pre-trained to alleviate the inaccurate embedding problem caused by the sparse interaction of target nodes;secondly,the designed multi-action coding module is used to learn the embedding of target nodes in the multi-action user-item space.The influence of complex dependencies between different behaviors on the embedding quality of target nodes is solved;finally,the multi-head self-attention mechanism is used to solve the social inconsistency problem of users in the social relationship network.The main research contents of this paper are as follows:In the pre-training process,in view of the delay effect brought by the inaccurate embedding of the target node’s cold-start neighbor nodes to the target node’s embedding,a neighbor meta-aggregator is proposed to enhance the information aggregation capability of each step of graph convolution,so that the model can The accurate embedding of the target node is predicted based on a few first-order neighbors and high-order neighbors.In order to ignore the influence of cold-start neighbors on the embedding of the target node,an adaptive neighbor sampler is proposed,which selects effective neighbor nodes according to the feedback of the graph neural network.In the message aggregation link,in view of the problem that existing algorithms ignore the dynamic nature of user-item interaction behavior,the interaction behavior time context information is integrated into the message propagation embedding,so that the model can learn the user interest transfer contained in the dynamic interaction behavior.Aiming at the complex dependencies between different user-item interaction behaviors,a multi-head attention-based interrelation encoder is proposed to learn the dependencies between different behaviors in pairs,thereby further improving the embedding quality of target nodes.When aggregating messages on a target node in a user social relationship graph,aggregating two neighbors with different interests at the same time will affect the embedding of the target node,and aggregating information from inconsistent social neighbor nodes will reduce the heterogeneous graph neural network.The problem of modeling the ability to use useful information in recommendation.Pre-MBRec captures consistent friend nodes in the user node information dissemination in the user-side social relationship graph through the multi-head self-attention mechanism.Experiments on Movie Lens and Yelp,a multi behavior dataset,show that,compared with five traditional recommendation models with better effects,such as NGCFM and NCFN,HR improves 7.03%～17.88%,and NDCG improves 16.67%～35.76%.In the personalized e-commerce recommendation task,compared with the five multi-action recommendation models and the recommendation model with auxiliary information such as MBGCN and KGAT,HR improves 10.8%～33.71% and NDCG improves 14.06% ～34.96%.Experiments verify the effectiveness of the Pre-MBRec model to deal with the above problems.