In brief
What the paper does.
Post-click conversion-rate estimation for recommender systems suffers from sample selection bias and data sparsity. The paper shows two defects of entire-space multi-task models, an inherent estimation bias and a potential independence priority, and proposes ESCM², a counterfactual risk minimizer that fixes both, with an analysis of its tuning and properties and an industrial deployment.
Cite
Cite the paper.
@article{wang2022escl,
title={Entire Space Counterfactual Learning: Tuning, Analytical Properties and Industrial Applications},
author={Wang, Hao and Chen, Zhichao and Fan, Jiajun and Huang, Yuxin and Liu, Weiming and Liu, Xinggao},
journal={arXiv preprint arXiv:2210.11039v2},
year={2022}
}