Dual cycle generative adversarial networks for web search

In this work, the IRGAN model is revisited to tackle semi-supervised information retrieval (IR) problems, considering the premature convergence of IRGAN caused by mismatching the guidance information between the generator G and the discriminator D. A novel framework called COGAN is proposed, which adds a tutor T to enhance G for avoiding the ill-trained models. The COGAN model plays a collaborative minimax game among T, G and D by iteratively optimizing them in two cycles of generative adversarial networks, including a tutorstudent cycle and a generator-discriminator cycle. COGAN has three compelling advantages over IRGAN: (i) alleviate the ill-trained problem, (ii) allow three models to play adversarial games, and (iii) prevent the model from overfitting. The COGAN framework is then applied in the web search task, and it shows significant improvement over other strong baselines. The improvements of COGAN over IRGAN in terms of NDCG@3 and MRR are 19.23% and 11.46%, respectively, on the largest semi-supervised learning to rank dataset MQ2008semi.