Non-parametric discriminatory power

Abstract

Discriminatory power is the relative usefulness of a feature for classification. Traditionally feature-selection techniques have defined discriminatory power in terms of a particular classifier. Non-parametric discriminately power allows feature selection to be based on the structure of the data rather than on the requirements of any one classifier. In previous research, we have defined a metric for non-parametric discriminatory power called relative feature importance (RFI). In this work, we explore the construction of RFI through closed-form analysis and experimentation. The behavior of RFI is also compared to traditional techniques.