[2D4-OS-7b] 広告とAI(2/2)

[2D4-OS-7b-04] Multi-task Delayed Feedback Model for Multiple Campaigns

〇Kanata Satake¹, Makoto Yamada^2,3, Shota Yasui⁴, Hisashi Kashima^2,3 (1. Graduate School of Informatics, Kyoto University, 2. Kyoto University, 3. RIKEN Center, 4. CyberAgent, Inc.)

Keywords:CVR Prediction, Delayed Conversion, Multi-task Learning

CVR prediction is important in online advertising because it strongly reflects the interests of users. Many studies have performed CVR prediction on datasets with multiple advertisements, which may lead to problems inherent in multi-task learning, such as negative transfer. In this study, we introduce multi-task learning to Delayed Feedback Model, which is a typical method for CVR prediction. The proposed method can deal with problems such as task size imbalance and negative transfer, and can optimize appropriately for each advertisement. Experiments on multiple datasets confirm the effectiveness of the proposed method.