Graph Neural Networks for Recommender Systems

In recent years, a new type of deep learning models, Graph Neural Networks (GNNs), have demonstrated to be a powerful learning paradigm when applied to problems that can be described via graph data, due to their natural ability to integrate representations across nodes that are connected via some topological structure. One of such domains is Recommendation Systems, the majority of whose data can be naturally represented via graphs. For example, typical item recommendation datasets can be represented via user-item bipartite graphs, social recommendation datasets by social networks, and so on. The successful application of GNNs to the field of recommendation, is demonstrated by the state of the art results achieved on various datasets, making GNNs extremely appealing in this domain, also from a commercial perspective. However, the introduction of graph layers and their associated sampling techniques significantly affects the nature of the calculations that need to be performed on GPUs, the main computational accelerator used nowadays: something that hasn't been investigated so far by any of the architectures in the recommendation literature. This thesis aims to fill this gap by conducting the first systematic empirical investigation of GNN-based architectures for recommender systems, focusing on their multi-GPU scalability and precision speed-up properties, when using different types of hardware.