Graph Coarsening Boosts Recommendation System Efficiency and Quality

Graph-based recommendation systems are widely used in industry but face scalability challenges with massive datasets. A new framework addresses this by employing graph coarsening, which systematically reduces graph size while preserving structural integrity. This technique aims to balance predictive quality with computational efficiency. The proposed two-stage diffusion framework integrates graph coarsening with multi-step label propagation. Initially, domain-specific heuristics aggregate nodes into meaningful communities, shrinking the graph. A first diffusion process, using either a Label Propagation Algorithm (LPA) or a Graph Neural Network (GNN), generates coarse-grained predictions on this reduced graph. Finally, a second LPA within subgraphs refines these predictions to produce specific recommendations for individual users. Applied to a real-world telecommunications dataset, this method achieved up to a 24% improvement in NDCG@5 over full-graph LPA, with a lightweight GNN further boosting performance by over 50% at the cost of increased training time. The research highlights an optimal balance between scalability, latency, and recommendation quality.