Learning From Heterogeneous Data Based on Social Interactions Over Graphs

Graph convolutional neural network is a deep learning model for graph-structured data. It has become a popular approach for recommendation system research due to its powerful feature extraction and characterization learning capabilities. As for rating prediction in recommendation systems, most existing models based on graph convolutional networks use heterogeneous interaction information between users and items but lack sufficient use of homogeneous interaction information in the user and item graphs, thus it leads to the degradation of recommendation accuracy performance. For this purpose, this paper proposes some methods for constructing homogeneous interaction graph models that can be combined with heterogeneous interaction graphs to fully aggregate the node similarity and edge link information in the graph so that node embedding representations based on graph data can be better learned through graph convolutional networks. Experimental results based on several recommendation datasets show that the proposed homogeneous interaction graph can help the recommendation model to better mine the potential feature information and reduce the prediction error of ratings.