ReducedGCN: Learning to Adapt Graph Convolution for Top-N Recommendation

Eungi Kim, Chanwoo Kim, Kwangeun Yeo, Jinri Kim, Yujin Jeon, Sewon Lee, Joonseok Lee

Abstract: Graph Convolutional Networks (GCNs) have gained prominence in recommendation systems, leveraging collaborative signals from high-order neighbors. However, they face a challenge known as over-smoothing, where user and item representations become excessively similar when the model gets deeper, contradicting the goal of personalized recommendations. The main cause of over-smoothing lies in prevailing message passing, which indiscriminately aggregates all neighbors using deterministic weights regardless of actual relevance. In this paper, we propose ReducedGCN, equipped with macro-scale Neighborhood Reduction and micro-scale Edge weight Reduction for less relevant interactions. They mitigate over-smoothing by selectively reducing message passing from irrelevant neighbors and edges, while maintaining that from distant but relevant ones. Our approach learns the degree of reduction for each individual node or edge completely from the data, distinguished from previous works using random sampling or heuristics. It is also generally applicable to various GCN models. Our comprehensive experiments verify its superior performance over the current state-of-the-art models.

External IDs:doi:10.1007/978-981-96-8180-8_23