A1141
Title: Forward variable selection for ultra-high dimensional models
Authors: Toshio Honda - Hitotsubashi University (Japan) [presenting]
Abstract: Forward variable selection procedures are described with stopping rules for feature screening in ultra-high dimensional quantile regression models and ultra-high dimensional generalized varying coefficient models. For such very large models, penalized methods like Lasso and SCAD do not work numerically, and some preliminary feature screening is necessary before such penalized methods are applied. The desirable theoretical properties of the forward procedures are presented by taking care of uniformity w.r.t. subsets of covariates properly. The necessity of such uniformity has been often overlooked in the literature. The stopping rules suitably incorporate the model size at each stage of the forward variable selection procedures. The results of numerical studies are also presented, and it is talked about possible extensions.
