Research On Hybrid Recommendation System Based On Graph Neural Network

Research on graph neural networks is increasingly being applied to collaborative filtering.The traditional method learns the feature embeddings representation of users(items)through ID and attribute configuration information.The state-of-the-art graph convolution model effectively learns the associations between nodes in the user-item bipartite graph,and obtains multiple embeddings with richer semantics.However,as the depth of GCN increased,repeated graph convolution operations will result in over smoothing of the high-order features of nodes(i.e.the embedding representation is too similar).In addition,the simple interaction function(the inner product and the Hadamard product)is used to infer a user's preferences.It only considers the shallow interaction of multiple embeddings in the corresponding dimension,and is not enough to capture the high-order relationships between all embedding representations.In response to the above problems,this paper mainly gives the following two aspects of innovative research:First,in view of the limitations of the previous graph convolutional message propagation,a graph collaborative filtering algorithm with decoupled message propagation and feature extraction is proposed.In graph convolution(GCN)propagation,feature extraction is performed separately in each hidden layer,and message propagation and node feature representation of graph convolution are separated.The interaction between high-order and loworder node representation is also integrated in the feature extraction process.It is helpful to alleviate the problem that nodes embeddings are too similar.In the graph structure,the node representation after iterative information dissemination not only includes the explicit interaction between user nodes and neighboring item nodes,but also captures the potential relationships between distant nodes.Second,a new recommendation algorithm model called ConvGCCF is proposed,which uses convolutional neural network(CNN)to extract multiple embeddings interactions in graph collaborative filtering.Specifically,the outer product is used to model the paired two sets of multiple embeddings to form an interaction cube with all dimensions of the embeddings.Use CNN to extract the multi-dimensional relationships and high-order interactive associations of interaction cube.The residual strategy is used to fuse the high-order and low-order interactions,and finally the user preference value is inferred.Finally,We have done a lot of experiments on the two public data sets Yelp and Movielens1m,which confirmed that our model has a great positive effect on the performance of the recommendation system.