Graph-based Recommendation using Graph Neural Networks

Graph based recommendation strategies are recently gaining momentum in connection with the availability of new Graph Neural Network (GNN) architectures. In fact, the interactions between users and products in a recommender system can be naturally expressed in terms of a bipartite graph, where nodes corresponding to users are connected with nodes corresponding to products trough edges representing a user action on the product (usually a purchase). GNNs can then be exploited and trained in order to predict the existence of a specific edge between unconnected users and product, highlighting the interest for a potential purchase of a given product by a given user. In the present paper, we will present an experimental analysis of different GNN architectures in the context of recommender systems. We analyze the impact of different kind of layers such as convolutional, attentional and message-passing, as well as the influence of different embedding size on the performance on the link prediction task. We will also examine the behavior of two of such architectures (the ones relying on the presence of node features) with respect to both transductive and inductive situations.