Estimating True Post-Click Conversion via Group-stratified Counterfactual Inference

Tiankai Gu, Kun Kuang, Hong Zhu, Jingjie Li, Zhenhua Dong, Wenjie Hu, Zhenguo Li, Xiuqiang He and Yue Liu

In online search and display advertising, the click-through rate (CTR) and the post-click conversion rate (CVR) are key measures of ad/campaign effectiveness. We find, however, both CTR and CVR are not always fair for advertisers to charge because of the "free-rider", referring to the user who inherently intended to make a conversion no matter with the ads promoting or not, but acted through promoted ads passingly. To tackle this problem, we propose a new measure, namely true post-click conversion rate (TCVR), to count the users who are truly affected by the ads promoting (i.e., users that made a conversion under ads, but no conversion if no ads.) under the Neyman-Rubin potential outcome framework. Theoretically, we demonstrate the advantages of our proposed TCVR for measuring ads effectiveness compared with the CTR and CVR. In the advertising scenarios, by assuming that all users can be stratified into five groups based on their behaviors with/without ads promoting under counterfactual overview, we can clearly identify the groups of users that are truly affected by the ads promoting. Moreover, to precisely estimate the TCVR, we propose an easy but effective counterfactual model, namely
Group-stratified Counterfactual Inference (GCI) algorithm, by counterfactually predicting the probability of each specific group of each unit belongs to. With empirical experiments, we demonstrate the effectiveness of our proposed counterfactual predictive model and confirm the advantages of our TCVR compared with CTR and CVR.