Research On Signed Network Recommender System Model Based On Graph Convolution

With the continuous development of information technology,the massive information provided by various Internet platforms makes people face the information overload problem.The recommender system exploit key information from massive information and provides personalized recommendation for users.Thanks to the recent development of graph convolutional neural networks(GCNs),recommendation algorithms can use this technique to model the robust representation for nodes by mining high-order similarities in the network,thus greatly improving the performance of the recommendation system.However,most recommender system models based on graph convolutional neural networks are based on unsigned networks,and they cannot be directly applied in the signed networks since the different link semantics.Considering that the semantics of links contained in signed networks can improve the quality of network node representation,this paper uses graph convolutional neural networks to construct a recommendation system model for signed networks.The specific work is as follows:(1)This thesis proposes a signed social recommendation model based on cross-validation to alleviate the problem of social inconsistency.Through the item-space user embedding,obtain the dynamic user preference similarity,and thus validate the positive and negative links in the original signed social network.After cross-validation,extract three more credible signed relationships,from which reconstruct the signed social network.By performing graph convolution operations in the reconstructed signed social network and user-item bipartite network,obtain more robust user and item representations and obtain predictions for missing scoring terms in the form of inner product.At the end of this chapter,comparative experiments and ablation experiments show that the model has a better recommendation performance.(2)This thesis proposes a recommendation model for implicit signed relationship based on the user-item bipartite network.By thresholding the score,reconstruct the original user-item bipartite network into an implicit signed network.Through the graph convolution operation in the implicit signed bipartite network,obtain the positive embedding,negative embedding and final embedding for nodes,and get the two sets of score prediction values by using the user’s positive and negative embedding and item embedding,through the dynamic combination of two sets of predicted values,then obtain the final prediction score.At the end of this chapter,comparative experiments and ablation experiments show that the model has a better recommendation performance.The experimental results of the above model demonstrate the effectiveness of the proposed model,which to some extent fills the gap in the combination of graph convolutional neural networks and signed network recommendation systems.