Mathematic Colloquium via Zoom

Sufficient Variable Selection and Dimension Reduction via Expected Conditional Hilbert-Schmidt Independence Criterion

Dr. Chenlu Ke

Virginia Commonwealth University

Abstract

In this talk, we first introduce a model-free sufficient variable screening procedure for ultrahigh dimensional data based on a newly developed independence measure. Compared with sure independence screening and its family, our approach inherits the power of the new measure and incorporates joint information between variables additionally to achieve sufficient variable screening. The advantages of our method are illustrated theoretically and numerically. We then introduce a novel sufficient dimension reduction approach using the same independence measure. An algorithm is developed to search dimension reduction directions using sequential quadratic programming. The method can be applied after our variable selection procedure to further extract information from data. A real data example is presented to demonstrate the joint use of the two methods.

Zoom ID: 967 4863 5043. Passcode upon request.

Free