Large-scale Causal Approaches to Debiasing Post-click Conversion Rate Estimation with Multi-task Learning

摘要: Post-click conversion rate (CVR) estimation is a critical task in e-commerce recommender systems. This deemed quite challenging under the industrial setting with two major issues: 1) selection bias caused by user self-selection, and 2) data sparsity due to rare click events. A successful typically has following sequential events: "exposure -> conversion". Conventional CVR estimators are trained space, but inference done entire exposure space. They fail account for causes of missing treat them as at random. Hence, their estimations highly likely deviate from real values large. In addition, issue can also handicap many which usually have large parameter spaces. In this paper, we propose principled, efficient effective estimation, namely, Multi-IPW Multi-DR. The proposed models approach causal perspective not our methods based on multi-task learning framework mitigate issue. Extensive experiments industrial-level datasets show that outperform state-of-the-art models.