Research Of Recommendation Technology Based On Graph Convolution Network And Fusing Item Attribute Information

With the rapid development of information technology,people can enjoy great convenience brought by the Internet technology,while have to face the problem of information overload caused by the massive amount of data on the Internet.The recommendation system is an effective tool to alleviate the problem of information overload,which can discover products or services that users might be interested in from massive data and then generating personalized recommendation.Collaborative Filtering(CF)is a very popular and effective method for recommendation.However,since it only considers users' historical interactions and makes recommendations based on their potential common preferences,it usually may suffer from the sparsity of user-item interactions and the cold start problem in practical application.In recent years,many recommendation methods based on deep learning have been proposed to address these limitations of CF-based methods.Among them,Graph Convolution Networks(GCNs)for recommendation could learn user and item representation by aggregating multi-hop neighbors' information over the user-item bipartite graph,which leads to significant accuracy improvements.Although GCN-based recommendation methods have achieved good performance,they still have some limitations: on the one hand,they usually only capture the information of user-item historical interactions without using additional side information to improve the data sparsity and cold start problem;on the other hand,they only use the observed user-item interactions as the supervised signal to train the model,and when the observed interactions are extremely sparse,the model cannot obtain sufficient supervised signal to learn high-quality user and item representations.In order to make up the deficiencies of GCN-based recommendation methods,we introduce the item attribute information and self-supervised learning technology on the basis of them.The major work of this thesis is as follows:(1)We propose a recommendation method based on graph convolution network and fusing item attribute information.First,we use graph convolution network that propagates user and item representation over the user-item bipartite graph and user-item attribute information bipartite graph to integrate neighbors' information into node representation.After that we use the weighted fusion of user and item representations that are obtained from two bipartite graphs as the final representation.At last,we conduct inner product based on final user and item representation to estimate the user's preference towards the target item.(2)We propose a recommendation method based on self-supervised graph learning and fusing item attribute information.At first,we take the recommendation method based on graph convolution network and fusing item attribute information as the main supervised learning task.Then we change the structure of user-item bipartite graph and user-item attribute information bipartite graph as the input data of model to get the node representations from different views,which are used for self-supervised contrastive learning and thus obtaining more accurate user and item representation.(3)We preform experiments and provide analysis for these two recommendation methods proposed in this thesis.The experimental results have shown that our methods outperform all other baselines on recall and normalized discount cumulative gain,which indicates our methods have better recommendation accuracy.