Applications of Graph Neural Networks to Large-Scale Recommender Systems: Some Results

Recent machine learning algorithms exploit relational information within a graph based data model. This paper considers one of the most promising supervised machine learning methods, the graph neural networks (GNN), for this class of problems. An application of the GNN approach to a recommender system problem revealed certain limitations of the GNN. A recommender system makes personalized product suggestions by extracting knowledge from previous user interactions. This is a specially interesting scenario because of the scale-free nature of graphs modelling user preference dependencies. We focused on the MovieLens dataset, which contains data collected from a popular recommender system on movies, and has been widely used as a benchmark problem for evaluating recently proposed approaches. This paper performs a deep analysis on the dataset to help discovery of some intriguing properties, and discusses problems and limitations encountered by GNN while facing this particular practical problem.