学术报告：Transfer learning with invariance structures across target and auxiliary sources

报告时间：6月10日（星期二）下午16:00-17:00

报告地点：沙河校区，学院1号楼102会议室

主持人：杨玥含 教授

报告人：夏思薇，成都理工大学，讲师

报告摘要：This paper studies the transfer learning problem under multiple source data with invariance structure among all sources, aiming to improve the estimation and prediction of the target source by borrowing knowledge from auxiliary sources. To boost transfer learning, we propose a novel estimation method called Joint Invariant Adaptive Transferred Estimation (JIATE), enabling a more accurate evaluation. JIATE learns the invariance structure by recognizing invariant features, which have similar effects on responses, in all sources, imposing adaptive penalty strengths to different features. It overcomes the challenge caused by the stringent similarity between target and auxiliary sources, and the noisy information from auxiliary samples. The theoretical analysis shows that JIATE guarantees consistency in selection and estimation. Extensive simulations and a real-data experiment on cancer data analysis verify the effectiveness and accuracy of our proposed method.

报告人简介：夏思薇，成都理工大学数学科学学院讲师。2021年博士毕业于重庆大学数学与统计学院。主要研究方向包括模型选择、特征选择、无模型方法和迁移学习。文章发表于Knowledge-Based Systems、Applied Mathematics and Computation等期刊。

撰稿人：刘洁

审稿人：邓露