Enhancing Recommendation Systems’ Performance with Network-Oriented Negative Sampling

ERDEM E. E., ORMAN G. K.

2024 Discovery Science Late Breaking Contributions, DS-LB 2024, Pisa, Italy, 14 - 16 October 2024, vol.3928, (Full Text) identifier

  • Publication Type: Conference Paper / Full Text
  • Volume: 3928
  • City: Pisa
  • Country: Italy
  • Keywords: Data Sparsity, Negative Sampling, Network Topology, Recommendation Systems
  • Galatasaray University Affiliated: Yes

Abstract

Recommendation systems often struggle with data sparsity and class imbalance, affecting their ability to suggest relevant items. An efficient solution to these problems is negative sampling. This study introduces three novel network-oriented negative sampling strategies—Path-Length (PL-NS), Diffusion Distance (DD-NS), and Path-Density (PD-NS)—to enhance the GCN-based recommendation systems’ performance. Integrated into the LightGCN framework and tested on the Last.fm dataset, our methods showed up to a 5% improvement in Precision@10, 6.7% in NDCG@10, and reduced training time by 30-35%, highlighting both performance and efficiency gains.