A Graph Neural Network Framework for Social Recommendations

Data in many real-world applications such as social networks, users shopping behaviors, and inter-item relationships can be represented as graphs. Graph Neural Networks (GNNs) have shown great success in learning meaningful representations for graphs by inherently integrating node information and topological structure. Data in social recommendations can also be denotes as graph data in the form of user-user social graphs and user-item graphs. In addition, the relationships between items can be denoted as item-item graphs. GNNs provide an unprecedented opportunity to advance social recommendations. However, there are tremendous challenges in building GNNs-based social recommendations where (1) users (items) are simultaneously involved in the user-item graph and user-user social graph (item-item graph); (2) user-item graphs not only contain user-item interactions but also include users’ opinions on items; and (3) the nature of social relations are heterogeneous among users. In this paper, we propose a novel graph neural network framework ( <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">GraphRec+</b> ) for social recommendations, which is able to coherently model graph data in order to learn better user and item representations. Specifically, we introduce a principled approach for jointly capturing interactions and opinions in the user-item graph and also propose an attention mechanism to differentiate the heterogeneous strengths of social relations. Comprehensive experiments on three real-world datasets show the effectiveness of the proposed framework.