GANRec: A negative sampling model with generative adversarial network for recommendation

• Published in: Expert systems with applications 2023-03, Vol.214, p.119155, Article 119155
• Main Authors: Yang, Zhi, Qin, Jiwei, Lin, Chuan, Chen, Yanping, Huang, Ruizhang, Qin, Yongbin
• Format: Article
• Language: English
• Subjects: Bipartite graph; Generative adversarial network; Knowledge graph; Negative sampling; Recommendation
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2022.119155

Although many negative sampling and generative adversarial network (GAN) strategies have been applied to recommendation, this frequently encounter problems of poor interpretability and performance of negative sampling. To address these issues, we construct a recommendation model named GANRec, which organically integrates negative sampling and GANs on a bipartite graph (BG) or knowledge graph (KG). The primary purpose of GANRec is to train a discriminator (i.e., a recommender) by generating high-quality negative samples that satisfy the perspective of graph theory, core idea of collaborative filtering, and distribution of positive interactions. The generator attempts to produce, distinguish, and select an informative and knowledge-aware negative item that can reflect the real needs of an user. Experiments on the Amazon-Book, Last-FM, and Yelp2018 datasets show that the recommendation performance of the proposed method not only exceeds the KG-enhanced models, but also is better than those of state-of-the-art sampling methods. Additionally, when no external knowledge is available, GANRec outperforms some BG-enhanced recommendation methods. Our codes are available in https://github.com/Yangzhi22/GANRec. •We analyze suitable negative items from three aspects, to yield informative samples.•We develop a recommender system through negative sampling and adversarial network.•Negative items generated by GANRec may reflect the real interests and needs of users.•GANRec has better performance and more intuitive interpretability than other methods.•We prove the advantages of GANRec by numerous experiments on six datasets.