Department of
presents
Dr. Helen Zhang
hzhang at stat dot ncsu dot edu
ABSTRACT
Penalized regression with thresholding-penalty has become a popular approach to variable selection in linear models. These methods shrink some coefficients to exactly zeros and are able to select important variables and estimate coefficients simultaneously. The first part of this talk will review a variety of popular shrinkage methods. Their geometric motivation, theoretical properties and finite sampling performance are summarized. In the second part of the talk, we will introduce a new class of penalized likelihood methods based on a LASSO-type penalty, which imposes a weighted penalty on coefficients according to their relative importance. The weights are chosen adaptively by data. Theoretical and computational properties of the new estimators are presented. Furthermore, we will discuss the challenge in choosing proper weights for high dimensional data. Finally, the performance of the proposed methods will be illustrated in various contexts including ordinary linear regression, generalized linear regression and survival data analysis.
Friday, December 1, 2006
3:35 - 4:35 pm
206 Cox Hall
Refreshments will be served on the second floor of Dabney Hall (left of Room 222) at 3:00 pm.