Personalized Ad Quality Bidding with MTML Causal Modeling and Constrained Optimization

Libin Liu, Fang Liu, Chen Fu, Fei Peng, Ethan Shao, Wei Guan, Kevin K. Chang

Keywords: Advertisement, Personalization, Optimization, Causal learning

Abstract: Ad quality plays a central role in ranking systems, promoting high-quality ads and demoting low-quality ones to enhance user experience and ultimately drive long-term value for both people and businesses. The quality of each ad is estimated by a value model, which computes a weighted sum of various quality predictions. Since different user cohorts exhibit heterogeneous sensitivities to the same ad, personalization aims to customize these weights to achieve a more efficient trade-off between ads performance and user engagement. In this paper, we propose a new personalization framework with two key innovations: 1) a multi-task multi-label (MTML) causal model that jointly predicts user sensitivities across multiple ad quality signals; and 2) a user sensitivity information aware and structural information aware optimization framework for learning more efficient scalar weights. With these improvements, our framework achieves a 0.5% increase in ads performance while maintaining neutral engagement, and delivers a 1.4x gain in efficiency compared to the current production system.

Submission Number: 10