Negative sampling strategy based on multi-hop neighbors for graph representation learning

Abstract

Contrastive learning (CL) has recently achieved significant success in the field of recommendation system. However, current studies mainly focus on obtaining high-quality positive samples and focus less on selecting negative samples. In existing recommendation system based on graph contrastive learning, most methods select negative samples by randomly selecting samples that have not interacted with the target node. Although random negative sampling is easy to implement and has wide applicability, it may lead to problems such as unbalanced data distribution and selection of false negative samples, which can degrade model performance. To address the above issues, we propose a novel negative sampling strategy called the Multi-hop Neighbors Negative Sampling method, named NSHN. Specifically, we select the information of 3-hop neighbors of each node as candidate negative samples. In addition, to reduce the impact of false negative noise on negative samples, we propose an adaptive denoising training strategy that adaptively prunes noise interactions during training. Experimental results demonstrate that our method performs well on four datasets and outperforms graph contrastive learning methods that use random negative sampling. The source code is available at: https://github.com/zhangkaiyu-zky/NSHN