Preference-corrected multimodal graph convolutional recommendation network

Users purchase goods to form a user-item interaction graph, where goods are usually displayed in multiple modes. The main role of the recommendation system is to obtain users' preferences after analyzing users' purchasing behavior. However, there are some deviations in the interaction between users and items. For example, users buy goods they are not interested in, which directly affects the user's preference analysis. Most existing models do not focus on analyzing and correcting user preference errors in the user-item relationship graph, which leads to errors. There is little work on the multimodal information of commodities, resulting in the loss of information. In this paper, a preference-corrected multimodal graph convolution recommendation network (PMGCRN) is proposed to provide multimodal recommendation services for users. First, a multichannel network is designed to obtain user preference information under different modes. Then, a positive attention mechanism is proposed to deal with the implicit noise edges that do not match users' interactions in the user-item graph to correct errors in user preferences. Additionally, the simplified graph convolution network acts on the structural information of the user-item bipartite graph as an additional channel to enhance the robustness of the model. Finally, a self-attention mechanism and layer-by-layer superposition are applied to obtain multilevel modal and structural information, respectively, so that valuable information can be obtained by the fusion matrix. Experiments show that our proposed PMGCRN outperforms other baselines on all three datasets, MovieLens, Amazon - Sports and Outdoors, and Douban.