LightGCN: Simplifying and Powering Graph Convolution Network for Recommendation

Graph Convolution Networks had become the new state-of-the-art for collaborative filtering, but the reasons for their effectiveness in recommendation were not well understood, and prior adaptations lacked thorough ablation of GCN components originally designed for graph classification. Through empirical study, the authors find that the two most common GCN designs—feature transformation and nonlinear activation—contribute little to collaborative filtering performance and even make training harder and degrade results. They therefore propose LightGCN, which includes only the most essential GCN component, neighborhood aggregation: it learns user and item embeddings by linearly propagating them on the user-item interaction graph and uses the weighted sum of embeddings from all layers as the final representation.

This simple, linear, and neat model is much easier to implement and train, and it delivers substantial gains—about 16.0% average relative improvement over Neural Graph Collaborative Filtering (NGCF), a state-of-the-art GCN-based recommender, under exactly the same experimental setting. The authors further provide analytical and empirical analyses justifying the rationality of the simplified design, and LightGCN became a widely adopted, strong baseline for graph-based recommendation.