GGDHSCL: A Graph Generative Diffusion With Hard Negative Sampling Contrastive Learning Recommendation Method
Xiaoyang Liu, Guiling Wen, Aijuan Wang, Chao Liu, Wei Wang, Pasquale De Meo
Abstract
Recommender Systems in real scenarios suffer from poor representation ability of user–item interaction graph caused by data sparsity and data noise. Most of the existing models have problems of instability and limited generation ability. This article proposes a novel recommendation method called graph generative diffusion with hard negative sampling contrastive learning recommendation method (GGDHSCL) to overcome the limitations above. First, the latent diffusion model (L-diffusion) and parametric topological noise reduction network (PTDNet) were introduced as view generators to improve the limited representation ability and mitigate noise. Second, dual-view contrastive learning was constructed to alleviate the limitations of high-quality data in the recommendation system and the problem of model collapse in the training process. Third, a hard negative sampling strategy was proposed to improve the self-supervised signal. We extensively compared our method with 14 popular baselines on four public datasets (Yelp, BeerAdvocate, Gowalla, and LastFM). Experiments show an improvement of recommendation quality (e.g., that on the BeerAdocate dataset, NDCG@40 is improved by 3.6% and Recall@40 is improved by 3.3% on BeerAdvocate dataset).