Model-Free Checking Meets Cross Domain Data: A Transfer Learning Approach

主讲人：许王莉 中国人民大学教授

时间：2026年9月23日16：00

地点：国际交流中心明远厅

举办单位：数理学院

主讲人介绍：中国人民大学吴玉章讲席教授，博士生导师。先后主持5项国家自然科学基金，北京市自然科学基金重点研究专题，教育部人文社会科学重点研究基地重大项目和教育部人文社科基金等多项科研课题。在顶尖期刊JASA, JRSSB, Biometrika, TPAMI等发表百余篇论文。先后获得中国第十二届北京市统计科研优秀成果奖一等奖(2014)，第一届统计科学技术进步二等奖（2021）。

内容介绍：High dimensional data are ubiquitous in many fields, yet limited sample sizes in target domains often compromise the power of statistical tests. Transfer learning provides a promising framework for leveraging data from source domains. However, existing methods, which are largely designed for estimation, rely heavily on stringent parametric assumptions or explicit structural similarities between domains. Such assumptions are difficult to verify in practice. Moreover, little research has been dedicated to exploring transfer learning specifically for hypothesis testing. To address this gap, we propose a novel model-free test that employs transfer learning via a projection-oriented method for high dimensional hypothesis testing. Our approach requires only a mild nonorthogonal similarity condition, which is naturally satisfied in most practical settings, thereby circumventing restrictive assumptions. The proposed statistic converges to a chi-squared distribution and is more efficient than a statistic constructed from target data alone. Notably, our method maintains its validity under relaxed conditions on covariate sparsity relative to conventional high dimensional inference.
The framework is extensible to multi-source settings and leads to improved power whenever at least one source dataset satisfies the nonorthogonal similarity condition. Extensive numerical studies demonstrate its superiority over methods without transfer learning.