Disentangled progressive negative sampling for graph collaborative filtering recommendation

doi:10.1016/j.knosys.2025.114133

Disentangled progressive negative sampling for graph collaborative filtering recommendation

2025-07-19

0

PRE

AI

DOI：10.1016/j.knosys.2025.114133

Original

Original request for help

Save

Abstract

En 中文

Graph Neural Networks (GNNs) are widely applied in collaborative filtering (CF) recommendations, learning user and item representations from large-scale interaction data using negative sampling. However, existing negative sampling methods for graph-based CF face two limitations: (1) underutilization of feedback signals due to GNNs’ message-passing mechanisms, and (2) the false negative problem caused by fixed sampling distributions. Graph neural networks (GNNs) are widely utilized in collaborative filtering (CF) recommendations to learn user and item representations from large-scale interaction data using negative sampling. However, existing negative-sampling methods for graph-based CF face two limitations: (1) underutilization of feedback signals owing to the message-passing mechanisms of GNNs, and (2) the false-negative problem due to fixed sampling distributions. Therefore, we propose a novel framework, disentangled progressive negative sampling (DPNS), for graph-based CF. DPNS introduces flexible sampling distributions and disentangles positive and negative feedback by focusing on wider ranks during negative sampling to avoid feedback conflict in low-rank spaces. It comprises three phases: (1) disentangling low- and wider-rank eigenvalue spaces via semantic contrastive learning and applying a wide-rank-aware sampling strategy to guide negative feedback into the wider-rank space, (2) generating synthetic hard negatives to optimize positive feedback in the low-rank space, and (3) integrating an adaptive gradient-reversal technique with Bayesian personalized ranking loss to reduce false negatives. Experiments on three real-world datasets demonstrated significant performance gains. For example, on the Amazon Patio dataset with LightGCN, DPNS improved recall (+24.4 %), NDCG (+18.9 %), and hit ratio (+20 %), evidencing its effectiveness in enhancing graph-based CF models. The code is available at https://github.com/wayneHallway/DPNS .

Keywords:

Graph Neural Networks

Collaborative Filtering

Negative Sampling

Disentangled Learning

Recommendation Systems

Disentangled progressive negative sampling for graph collaborative filtering recommendation

Abstract

Journal

Organization