Sequential Lasso Cum EBIC for Feature Selection With Ultra-High Dimensional Feature Space

Abstract

In this article, we propose a method called sequential Lasso (SLasso) for feature selection in sparse high-dimensional linear models. The SLasso selects features by sequentially solving partially penalized least squares problems where the features selected in earlier steps are not penalized. The SLasso uses extended BIC (EBIC) as the stopping rule. The procedure stops when EBIC reaches a minimum. The asymptotic properties of SLasso are considered when the dimension of the feature space is ultra high and the number of relevant feature diverges. We show that, with probability converging to 1, the SLasso first selects all the relevant features before any irrelevant features can be selected, and that the EBIC decreases until it attains the minimum at the model consisting of exactly all the relevant features and then begins to increase. These results establish the selection consistency of SLasso. The SLasso estimators of the final model are ordinary least squares estimators. The selection consistency implies the oracle property of SLasso. The asymptotic distribution of the SLasso estimators with diverging number of relevant features is provided. The SLasso is compared with other methods by simulation studies, which demonstrates that SLasso is a desirable approach having an edge over the other methods. The SLasso together with the other methods are applied to a microarray data for mapping disease genes. Supplementary materials for this article are available online.