Scalable Graph Neural Network Training via Distributed Subgraph Sampling for Enhanced Recommendation Systems

Graph Neural Networks (GNNs) have demonstrated significant potential in enhancing the performance of recommender systems. However, training GNNs on large-scale recommendation graphs presents substantial computational challenges due to memory limitations and communication overheads. This paper proposes a distributed subgraph sampling strategy to address these challenges, enabling efficient and scalable GNN training. Our approach leverages a distributed computing framework to partition the recommendation graph and independently sample subgraphs on each worker node. These subgraphs are then used to train a GNN model using a distributed stochastic gradient descent algorithm. We demonstrate the effectiveness of our approach through experimental evaluation on a synthetic dataset mimicking real-world recommendation graphs, showing significant improvements in training time and scalability compared to centralized training methods. The implications of this research are particularly relevant in light of recent advancements in representation learning for large-scale recommender systems, as discussed by Gaggi (2026).