An in-depth study on adversarial learning-to-rank

In light of recent advances in adversarial learning, there has been strong and continuing interest in exploring how to perform adversarial learning-to-rank. The previous adversarial ranking methods [e.g., IRGAN by Wang et al. (IRGAN: a minimax game for unifying generative and discriminative information retrieval models. Proceedings of the 40th SIGIR pp. 515-524, 2017)] mainly follow the generative adversarial networks (GAN) framework (Goodfellow et al. in Generative adversarial nets. Proceedings of NeurIPS pp. 2672-2680, 2014), and focus on either pointwise or pairwise optimization based on the rule-based adversarial sampling. Unfortunately, there are still many open problems. For example, how to perform listwise adversarial learning-to-rank has not been explored. Furthermore, GAN has many variants, such as f-GAN (Nowozin et al. in Proceedings of the 30th international conference on neural information processing systems, pp. 271-279, 2016) and EBGAN (Zhao et al. in Energy-based generative adversarial network. International conference on learning representations (ICLR), 2017), a natural question arises then: to what extent does the adversarial learning strategy affect the ranking performance? To cope with these problems, firstly, we show how to perform adversarial learning-to-rank in a listwise manner by following the GAN framework. Secondly, we investigate the effects of using a different adversarial learning framework, namely f-GAN. Specifically, a new general adversarial learning-to-rank framework via variational divergence minimization is proposed (referred to as IRf-GAN). Furthermore, we show how to perform pointwise, pairwise and listwise adversarial learning-to-rank within the same framework of IRf-GAN. In order to clearly understand the pros and cons of adversarial learning-to-rank, we conduct a series of experiments using multiple benchmark collections. The experimental results demonstrate that: (1) Thanks to the flexibility of being able to use different divergence functions, IRf-GAN-pair shows significantly better performance than adversarial learning-to-rank methods based on the IRGAN framework. This reveals that the learning strategy significantly affects the adversarial ranking performance. (2) An in-depth comparison with conventional ranking methods shows that although the adversarial learning-to-rank models can achieve comparable performance as conventional methods based on neural networks, they are still inferior to LambdaMART by a large margin. In particular, we pinpoint that the weakness of adversarial learning-to-rank is largely attributable to the gradient estimation based on sampled rankings which significantly diverge from ideal rankings. Careful examination of this weakness is highly recommended for developing adversarial learning-to-rank approaches.