A Novel Nonparametric Variable Selection for Ultra High-Dimensional Problems

• Activity Number: 303
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Subhadeep Mukhopadhyay*+
• Companies: Texas A & M University
• Keywords: high dimension ; classification ; comparison density ; mid-distribution score function

Abstract

In contrary to the standard practice of variable selection using penalized likelihood method, in this paper we introduce a fast Non-parametric classifier-independent feature selection algorithm for high dimensional classification problem using the concept of comparison density, mid-distribution score function. The main benefits of our proposed method (i) robust nonparametric model-free, (ii) can detect non-linear relationship between response and covariates, (iii) can handle continuous, discrete as well as categorical predictor variables, and (iv) can give a deeper insight on the questions 'why`` and 'when`` as well as 'how`` a variable could be influential through visualization. We find that efficiency and effectiveness of our methods and offers additional interpretability and understanding into the operation of variables. Yet surprisingly all available methods are silent on this point. This extra piece of information can help applied researchers not only to identify influential variables but to identify the distinctive characteristics of the variable which makes it so useful for classification - an insightful data analytic variable selection.