Abstract
Product recommendation systems are fundamental to modern e-commerce platforms, yet existing approaches face significant challenges with data sparsity, cold-start problems, and scalability limitations in large-scale applications. To address these challenges, this paper introduces GNNRec, a graph-based neural network model for efficient and robust product recommendation. It is a comprehensive graph-based recommendation framework that extends the LightGCN architecture with multi-relational user-item interactions, heterogeneous node embeddings, and multi-task learning capabilities. Our model incorporates rich semantic features and edge-weighted graph propagation to effectively capture complex user preferences and item characteristics. GNNRec addresses the cold-start problem through sophisticated graph propagation techniques and content-based initialization strategies, and handles data sparsity via adaptive sampling strategies. Extensive experiments on large-scale real-world datasets demonstrate that GNNRec achieves significant improvements over state-of-the-art baselines, achieving an AUC of 0.916 and consistent superior performance across all ranking metrics. Notably, GNNRec achieves NDCG@10 of 0.0428, representing a 4.4% improvement over the strongest baseline (NGCF at 0.0410). Moreover, GNNRec demonstrates strong robustness under data sparsity (99.988%), cold-start conditions (16.3% improvement for users with 5-10 interactions, and viable recommendations for zero-interaction users), and temporal dynamics, while achieving 42.4% faster training with 33.3% lower memory usage, confirming its scalability and effectiveness for practical deployment.
Funding Information
  • BK21 FOUR R&E Center for Computer Science and Engineering, Korea University
  • Korea University
  • Tech Icubation Program for Startup (TIPS) program (RS-2025-25455127)
  • Seoul Metropolitan City's Regional Innovation System & Education (NARS3270Q)

This publication has 32 references indexed in Scilit: