Ecgan-LTR

Existing approaches of learning to rank based on generative adversarial networks (GANs) tend to suffer from different training problems of traditional GANs, such as training instability and mode collapse. Meanwhile, using fixed and static single adversarial object defined in GANs and its variants to iteratively and alternately train generators and/or discriminators is not conducive to improving the generative and/or discriminative performance from multiple perspectives, thus affecting the effectiveness of adversarial learning to rank methods, which still needs to be further improved. To address the aforementioned weakness, a novel listwise method of learning to rank is proposed, named ECGAN-LTR, which is the first attempt of the learning to rank method by integrating conditional GANs with evolutionary algorithm in the article. In the ECGAN-LTR method, evolutionary operations such as variation, evaluation, and selection are integrated into the framework of conditional GANs to simultaneously evolve the generator population and the discriminator population, and the adversarial game is played with the generator-discriminator pairs to alternately and iteratively train the optimal models of the generators and the discriminators. In each evolution of ECGAN-LTR, the generator adapts to the current environment (i.e. discriminator) according to multiple different loss functions (including a rich family of f-divergence functions) and these individuals with higher fitness will be retained. Extensive experiments are conducted on multiple public benchmark datasets of learning to rank, and demonstrate that our proposed ECGAN-LTR method achieves competitive and superior performance, and the overall results show that the ECGAN-LTR approach outperforms other current state-of-the-art adversarial learning to rank methods in the vast majority of cases

Adversarial system · Discriminative model · Discriminator · Evolutionary algorithm · Generative grammar · Adversarial Robustness in Machine Learning · Domain Adaptation and Few-Shot Learning · Generative Adversarial Networks and Image Synthesis