Graph neural networks for recommendation systems: A survey and future directions
Abstract
Graph Neural Networks (GNNs) are now a powerful means of learning about data in a graph format, with special support of recommendation systems, where the interactions between users and items naturally form complex graphs. This survey provides a close-up view of the application of GNNs in recommendation systems. By highlighting their contribution to representation learning through the acquisition of context, user-item relations, and higher-order relations. In a broad sense, we categorize existing GNN-based recommendation models into four primary categories that are knowledge-aware recommendation, session-based recommendation, user-item bipartite graph methods, and social recommendation. The significant aspects of design, like graph building, aggregation methods, and scalability concerns, are discussed in detail. We also analyze the GNNs' performances in various areas and benchmarks, which are more advantageous than traditional deep learning and collaborative filtering methods. GNN-based recommenders have drawbacks in terms of explainability, dynamic modelling, and scalability, notwithstanding their potential. We list potential research avenues to solve issues, such as effective large-scale training, including temporal and heterogeneous data, and making recommendations that are interpretable. For scholars and practitioners looking to comprehend and progress the state-of-the-art in GNN-powered recommendation systems, this survey attempts to act as a fundamental resource.