HaGAR: Hardness-aware Generative Adversarial Recommender

Implicit Collaborative filtering is a fundamental technique in recommendation systems, leveraging implicit user interactions to suggest items of interest. A significant challenge in this domain is the absence of explicit negative feedback, limiting the recommendation performance. Previous researchers have tried to tackle the challenge through the Generative Adversarial Network (GAN). The generator produces increasingly challenging samples for the discriminator, driving the optimization of the discrimination objective. Although GAN-style recommender systems can achieve decent performance by generating harder negative samples, the negatives selected by the generator may not always be ideal for training the discriminator. In this study, we focus on two types of undesirable negatives that persist in modern GAN-style recommenders: false negatives and uninformative negatives. In response to these issues, we propose a novel Hardness-aware Generative Adversarial Recommender (HaGAR). To the best of our knowledge, it is the first adversarial recommender that explicitly aims to alleviate the adverse impact of false and uninformative negatives. Our approach incorporates a relevance monitoring module and a hardness-aware weighting module to identify and address false and uninformative negatives during training with minimal additional computational cost. Our experimental results demonstrate that HaGAR significantly improves recommendation performance, achieving over a 21% increase in terms of NDCG@10 compared to the state-of-the-art GAN-style recommender. These findings highlight the efficacy of our improvement in providing more robust negative samples, leading to better-performing recommendation systems.