A Novel Framework for Robust Machine Learning–Based CATE Estimation

Abstract

In this talk, I will present our recent progress on Conditional Average Treatment Effect (CATE) estimation. I will begin by discussing the robustness issues we observed in industrial practice, where CATE estimation relies on machine learning (ML) techniques to handle high-dimensional data. In contrast to prior works focusing on empirical studies of the ML components, we propose a novel theoretical framework that reveals the ML-wise and CATE-wise bottlenecks in estimation. I will discuss how this framework addresses the robustness challenges and serves as a guidance for designing a more effective model training paradigm. Finally, I will demonstrate a simple training approach following the framework and is able to significantly improve estimation performance on both well-known benchmarks and real-world industrial data.