APL: Adversarial Pairwise Learning for Recommender Systems. (15th March 2019)

• Record Type: Journal Article
• Authors: Sun, Zhongchuan; Wu, Bin; Wu, Yunpeng; Ye, Yangdong
• Is Part Of: Expert systems with applications. Volume 118(2019)
• Page Start: 573
• Page End: 584
• Publication Date: 2019-03-15
• Subjects: Adversarial learning -- Pairwise ranking -- Matrix factorization -- Recommender systems; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2018.10.024
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: An adversarial pairwise learning model named APL is proposed for recommender systems. A differentiable procedure is adopted to replace the discrete item sampling. Three pairwise loss functions are evaluated under multiple recommendation scenarios. APL considerably improves the stability and convergence of adversarial learning. Abstract: The main objective of recommender systems is to help users select their desired items, where a major challenge is modeling users' preferences based on their historical feedback (e.g., clicks, purchases or check-ins). Recently, several recommendation models have utilized the adversarial technique, which has been successfully used to capture real data distributions in various domains (e.g., computer vision). Nevertheless, the training process of the original adversarial technique is very slow and unstable in the domain of recommender systems. First, the sparsity of the implicit feedback dataset aggravates the inherently intractable adversarial training process. Second, since the original adversarial model is designed for differentiable values (e.g., images), the discrete items also increase the training difficulty. To cope with these issues, we propose a novel method named Adversarial Pairwise Learning (APL), which unifies generative and discriminative models via adversarial learning. Specifically, based on the weaker assumption that the user prefers observed items over generated items, APL exploits pairwise ranking to accelerate the convergence and enhance the stability of adversarial learning. Additionally, a differentiable procedure is adopted to replace the discrete item sampling to optimize APL via backpropagation and stabilize the training process. Extensive experiments under multiple recommendation scenarios demonstrate APL's effectiveness, fast convergence and stability. Our implementation of APL is available at: https://github.com/ZhongchuanSun/APL .