Learning Disentangled Representations for Ads Ranking

Xuxing Chen, Yan Xie, Jin Fang

Keywords: causal learning, recommendation systems, ads ranking

Abstract: Online advertising systems rely on ads recommendation to deliver personalized experiences and high-quality ad rankings. To achieve this, large-scale deep learning models are employed to process user-ad interactions for accurate user behavior predictions. However, existing approaches often struggle to maintain optimal performance due to the inherent complexities of data distribution shifts in online serving environments and the heterogeneity of user preferences. To address these challenges, we propose novel causality-aware group learning algorithms that generate disentangled representations to improve ads ranking performance. Our approach focuses on identifying fine-grained segments and specific defects in existing ads ranking models, and developing targeted model architectural or algorithmic patches to mitigate these limitations. Through extensive experiments, we demonstrate the benefits of our approach, showcasing its potential to enhance ads recommendation in modern-scale advertising systems.

Submission Number: 11