A0868
Title: Outlier robust feature screening via penalized marginal regression models
Authors: Yuta Umezu - Nagasaki University (Japan) [presenting]
Abstract: For handling ultra high-dimensionality, marginal screening now become standard tool for selecting important covariates relevant to response variable. A lot of recent developments of the methods are based on some generalization of the estimates of either marginal regression coefficients or relevant measures, such as Pearson correlation of coefficients. The former is known as a model-based method, and the latter is known as a model-free method. The purpose is to select covariates based on penalized marginal regression models, which is some extension of the one developed by a past study in the sense that their method is included in the current as a special case. Precisely, the selected covariates obtained by the referred study are equivalent to the non-zero elements of $\ell_1$-type penalized least squares estimates. Based on these estimates, the model is slightly generalized to take into account heavy tail noise. As a result, the proposed method is robust to outliers with a smaller computational cost than the existing ones. In addition, the effectiveness of the proposed method is presented through some simulation studies.
