A0540
Title: Feature selection for high-dimensional heteroscedastic regression models
Authors: PoHsiang Peng - National Tsing Hua University (Taiwan) [presenting]
HaiTang Chiou - National Chung Cheng University (Taiwan)
Hsueh-Han Huang - Academia Sinica (Taiwan)
Ching-Kang Ing - National Tsing Hua University (Taiwan)
Abstract: Feature selection for high-dimensional linear heteroscedastic models is considered. Inspired by the connection between the linear heteroscedastic function and the interaction model, a two-stage algorithm is designed to choose the relevant features in the aforementioned high-dimensional model. Moreover, when it is unknown whether the functional form of heteroscedasticity is linear or multiplicative, a data-driven method is provided to select between the two alternatives. The selection consistency of the proposed method is proved and its performance is illustrated via numerical simulations. Further, the method is applied to identify defective tools in the semiconductor manufacturing process.
