Personalized treatment rules account for individual patient characteristics and offer substantial potential to enhance clinical outcomes. However, existing approaches for learning optimal treatment rules often rely on strong assumptions regarding outcome regression and propensity score models. These assumptions may lead to biased estimates if the models are misspecified. Furthermore, to address uncertainty in outcomes, risk-aware objective functions are often considered. These objective functions depend on the entire outcome distribution, which presents challenges for estimation using traditional regression-based methods. To overcome these challenges, we propose CG-learning, a generative approach for estimating risk-aware optimal treatment rules. By employing a nonparametric model for the conditional outcome distribution, this method avoids restrictive structural assumptions. We further incorporate representation learning into CG-learning to handle high-dimensional unstructured covariates, and refer to the resulting method as UCG-learning. Theoretical results are established for the estimated treatment rules from both CG-learning and UCG-learning, including nonasymptotic bounds on regret and misassignment probabilities. Comprehensive simulations and real-data applications demonstrate that the proposed methods consistently outperform benchmarks.

报告人简介：

胡翔斌，现任职于北京理工大学成人直播 。2021年于香港理工大学获得博士学位，先后在新加坡国立大学和香港理工大学从事博士后研究工作，并于2025年入职北京理工大学，获批2026年国家级青年人才项目。主要研究方向包括生存分析与深度学习理论，重点关注复杂结构下的生存数据建模、非参数与半参数统计方法、以及深度学习模型的统计理论基础等问题。相关研究成果发表于JASA、Biometrika等期刊。