Graph Convolutional Networks with Dynamic Subgraph Sampling for Recommendation Systems

Graph neural networks provide a powerful frame-work for recommendation systems by capturing dependencies through message passing and modeling user-item interactions using an adjacency matrix. However, as the number of users and items increases, the adj acency matrix grows larger and sparser, leading to an over-reliance on the existing graph structure. These challenges require frequent retraining, which ultimately restricts scalability and diminishes recommendation performance. This pa-per presents a recommendation algorithm of dynamic sub-graph sampling to address the challenges of data sparsity and graph de-pendency in recommendation systems based on graph neural net-works. The proposed algorithm alleviates data sparsity and graph dependency issues by dynamically sampling subgraphs to estab-lish adaptive user-item interactions by shifting from traditional whole-graph training to subgraph-based training. The proposed approach enhances aggregation and sampling techniques, proving the efficacy of separate subgraph sampling for user and item nodes. Its effectiveness is validated through comparative, ablation, and hyperparameter experiments across multiple benchmark datasets, outperforming baselines consistently.