Context-Aware Modeling via Simulated Exposure Page for CTR Prediction

Xiang Li, Shuwei Chen, Jian Dong, Jin Zhang, Yongkang Wang, Xingxing Wang, Dong Wang

Abstract: Click-through rate (CTR) prediction plays a crucial role in industrial recommendation and advertising systems, which generate and expose multiple items for each user request. Although the user's click action on an item will be affected by the other exposed items (called contextual items), current CTR prediction methods do not exploit this context because CTR prediction is performed before the contextual items are generated. This paper introduces a solution Contextual Items Simulation and Modeling (CISM) to tackle this limitation. Specifically, we propose a near-line Context Simulation Center to simulate exposure page without affecting online service latency, and an online Context Modeling Transformer to learn user-wise context from the simulated results w.r.t. the candidate item. In addition, knowledge distillation is introduced to further improve CTR prediction. Extensive experiments on both public and industrial datasets demonstrate the effectiveness of CISM. Currently, CISM has been deployed in the online display advertising system of Meituan Waimai, serving the main traffic.