机器学习与数据科学博士生系列论坛（第七十一期）—— Transportability without the Mean Exchangeability Assumption
报告人：张宇航（北京大学）
时间：2024-05-09 16:00-17:00
地点：腾讯会议 446-9399-1513
摘要：
Randomized controlled trials are considered the gold standard of causal inference. However, blindly applying the results from the source populations to the target population without adjusting for their discrepancy will lead to erroneous conclusions. Most previous works rely critically on the mean exchangeability assumption which assumes that source populations and the target population share the same potential outcome expectation. However, as pointed out by many current studies, the mean exchangeability assumption is controversial and may be violated in many situations. When the mean exchangeability assumption is violated, the heterogeneity between the populations will lead to site selection bias which prevents us from applying classical transportability methods.
In this talk, we will introduce the synthetic treatment group estimator to transport the information from multiple randomized trials to a target population without the mean exchangeability assumption. By exploring the connection between the transportability problem and the synthetic control method, we propose two new identification assumptions complementary to the mean exchangeability assumption. We will also demonstrate the robustness of the synthetic treatment group estimator with respect to data contamination.
论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。
