Authors: Alisha Raju Shaikh, Prof. S. K. Said, Prof. M. C. Kolse

Abstract: Recommendation systems help users discover products, films and articles on digital platforms. Classical collaborative and content-based filtering struggle to capture non-linear relationships and high-order connectivity between users and items, especially when interaction data are sparse. Graph Neural Networks (GNNs) address this by modelling interactions as a graph in which users and items are nodes and interactions or attributes are edges, and by learning embeddings through message passing over neighbourhoods. This paper reviews GNN-based recommendation. It introduces graph convolution, graph attention and inductive sampling, then surveys landmark recommenders including graph convolutional matrix completion, PinSage, Neural Graph Collaborative Filtering, LightGCN, knowledge-graph attention networks and session-based GNNs. Training objectives such as Bayesian personalized ranking and evaluation with recall and normalized discounted cumulative gain are explained. A layered architecture covering graph construction, embedding propagation, prediction and serving is proposed, together with research gaps, objectives and a comparison of models. Challenges of over-smoothing, scalability, cold start, fairness and explainability are discussed. The review concludes that GNNs provide a principled way to exploit relational structure in recommendation, with simpler propagation often outperforming heavier designs.

DOI: https://doi.org/10.5281/zenodo.24099884