A Two-Stage Recommendation Framework with Graph Convolutional Networks

Graph neural networks (GNNs) are widely used in recommender systems. Over-smoothing degrades the recommendation performance for the node representation diversity is reduced. It is difficult to capture complex graph structures and patterns. To address this challenge, this paper proposes a two-stage recommendation framework. In the first stage, a linear residual graph convolutional network filters out high-order noise to enhance embedding representations while capturing high-order interactions between users and items. In the second stage, matrix completion recommendation leverages the user-item rating interaction graph to avoid over-smoothing via rating prediction. Additionally, a sample removal strategy is utilized to prevent label leakage for performance improvement. The proposed approach is extensively evaluated through comparative, ablation, and hyperparameter experiments. It outperforms compared methods across multiple benchmark datasets.