Adversarial-Enhanced Causal Multi-Task Framework for Debiasing Post-Click Conversion Rate Estimation
Xinyue Zhang,Cong Huang,Kun Zheng,Hongzu Su,Tianxu Ji,Wei Wang,Hongkai Qi,Jingjing Li
DOI: https://doi.org/10.1145/3589334.3645379
2024-01-01
Abstract:In real-world industrial scenarios, post-click conversion rate (CVR) prediction models are trained offline based on click events and subsequently applied online to both clicked and unclicked events. Unfortunately, unclicked events are inevitably difficult to estimate due to user self-selection, which leads to a degradation of CVR prediction accuracy. In order to estimate the prediction of unclicked events, the current mainstream Doubly Robust (DR) estimators introduce the concept of imputed errors. However, inaccuracies in imputed errors can increase the uncertainty in the generalization bound of CVR predictions, consequently resulting in a decline in the CVR prediction accuracy. To challenge this issue, we first present a theoretical analysis of the bias and variance inherent in DR estimators and then introduce a novel causal estimator that seeks to strike a balance between bias and variance within the DR framework, thus optimizing the learning of the imputation model in a more robust manner. Additionally, drawing inspiration from adversarial learning techniques, we propose a novel dual adversarial component, which learns from both the space level and the task level to eliminate the causal influence of input features on the CTR task (i.e., the click propensity), with the goal of achieving unbiased estimations. Our extensive experimental evaluations, conducted on both the widely used benchmark and the real-world large-scale Internet giant platform, convincingly demonstrate the effectiveness of our proposed scheme. Besides, we have released a high-quality industrial dataset named Tenc-UnionAds used for selection bias research in the advertising field.