报告题目：多值处理下个体处理效应的共形推断方法

报告人： 赵慧 教授 （中南财经政法大学）

报告时间：2026年11月22日（周日）上午11：30-12：30

报告地点：学院会议室106

报告摘要: Conformal prediction offers a distribution-free, flexible framework for uncertainty quantification. Yet existing conformal methods for individual treatment effects are largely confined to binary treatments. Extending them to multivalued treatments raises two difficulties: propensity-score-based weights are often statistically unstable, and contiguous prediction intervals necessarily cover the low-density regions separating the modes of a multimodal potential-outcome distribution, leading to inefficient prediction sets. We address both issues by constructing adaptive prediction sets for counterfactual outcomes under multivalued treatments, combining KuLSIF-based density-ratio estimation with highest predictive density (HPD) conformal scores computed from conditional density estimates. Both estimators share a reproducing kernel Hilbert space least-squares formulation, which simplifies theoretical derivation and analysis. The resulting prediction sets are robust to heteroscedasticity and multimodal outcome distributions and effectively eliminate low-density regions between distribution modes.

报告人简介：赵慧，中南财经政法大学统计与数学学院教授、博士生导师，2005年北京大学数理统计专业博士毕业，之后在中国科学院系统科学研究所与美国密苏里大学从事博士后研究。主持国家自然科学基金项目4项，编著研究生教材《高等统计学》。曾获湖北省高等学校教学成果奖二等奖，主持建设国家级一流课程《时间序列分析》。在JASA、Biometrics、SIM、SJS、《中国科学：数学》等国内外知名学术期刊发表论文近七十篇，研究兴趣覆盖高维统计分析、生存分析、纵向数据分析、图模型与因果推断等领域。