Academy of Mathematics and Systems Science, CAS Colloquia & Seminars
Speaker:
张军舰 教授, 广西师范大学
Inviter:
Title:
Variable Selection Methods in Change-point Detection
Time & Venue:
2017.5.22 9:00-10:00 S309
Abstract:
In this talk, we discuss some change-point detection problem by using variable selection technique. Especially we propose a new model named mixture-regression model which combines auto-regression and regression. We investigate the estimation of the location of change-points in mixture- regression model. Using a penalized least-square criterion with a l1-type penalty and some suitable modification, the problem is cast as a sparse regression, and can be solved by resorting to the least-absolute shrinkage and selection operator(Lasso). With the way to select the tuning parameter by minimizing information criterion, an efficient block-coordinate descent algorithm is developed to implement the novel method. With probability tending to one, this method can efficiently determine the unknown number of change-points, and the estimated break locations are sufficently close to the true break locations. Finally, numerical tests using synthetic and real data are presented, which corroborate the merits of the method in mixture-regression model.