SSM-MTO: A Causal Framework for Session-level Ads Load Optimization

Hui Chen, Patrick R. Johnstone, Taihui Li, Shu Wang, Chao Cen, Qinqin Zhu, Jizhe Zhang

Keywords: Ads Load Optimization, Multiple Treatment Optimization, Causal Learning, Debiased Data, A/B test

Abstract: Session-level dynamic ads load optimization seeks to strike a balance between user experience and ads performance score by delivering the right number of ads during online sessions on social networks and e-commerce platforms. Previous approaches struggle with several challenges, including treatment-induced bias, carry-over effects, and business constraints on the trade-off between ads performance score and engagement across different cohorts. To overcome those challenges, we propose to train a session sensitivity model (SSM) as a lever to adjust ads load for each session. Then, we adopt a Multi-Treatment Optimization (MTO) framework by incorporating business constraints to dynamically determine the optimal ads load for each session. The SSM is trained on the data collected from the debiased data collection experiment which randomizes the ads load at the session level to remove the confounding bias caused by ads load treatments. From the offline training data, we showed that the SSM-MTO identifies the efficient sessions for ads load treatment. Furthermore, the SSM-MTO has been put into online A/B tests to serve the online traffic which achieved better efficiency and better trade-off between ads performance score and user experience.

Submission Number: 6