An Entire Space Multi-gate Mixture-of-Experts Model for Recommender Systems

With the development of e-commerce, both advertisers and platforms pay more and more attention to the effectiveness of ads recommendation. In recent years, deep learning approaches with a mulit-task learning framework have shown to be effective in such recommendation systems. One main goal of these systems is to estimate the post-click conversion rate(CVR) accurately. However, higher click-through rate(CTR) for a product does not always lead to higher conversion rate(CVR) due to many reasons (e.g. lower rating). In addition, the overall performance of the recommendation system may not be optimal, since the usage of multi-task models (the CTR and CVR tasks) is often sensitive to the relationships of the tasks. In this paper, we propose a deep neural model under the Mixture-of-Experts framework (MoE), call ES-MMOE, in which a sub-network is used to promote samples with high CVR. The model can also be trained with the entire space by taking advantage of the Entire Space Multi-task Model (ESMM) model. Extensive experiments on a large-scale dataset gathered from traffic logs of Taobao’s recommender system demonstrate that ES-MMOE outperforms a number of the state-of-the-art models, including ESMM, with a relatively large margin.