Recommender Systems Enhanced by Graph Neural Networks

Graph neural networks, with their ability to capture relations well, are a modern and attractive field of research. The thesis analyses existing recommendation systems that exploit graph neural networks, namely SR-GNN and TGN, to capture complex patterns in user behaviour. Because the long-term user's action history is often unavailable, the work focuses on models that base their recommendations on the current session. The work compares the models against competitive k-nearest-neighbour baselines and outlines the strengths and weaknesses of both approaches. Part of the work is a review of the theoretical background of the analysed models and the related concepts from both the recommendation systems and graph neural networks fields.