Effective classification using feature selection and fuzzy integration

  • Authors:

  • Nick J. Pizzi;Witold Pedrycz

  • Affiliations:

  • National Research Council of Canada, Institute for Biodiagnostics, 435 Ellica Avenue, Winnipeg, MB R3B 1Y6, Canada and Department of Computer Science, University of Manitoba, Winnipeg, MB R3T 2N2, ...;Department of Electrical and Computer Engineering, University of Alberta, Edmonton, AB T6G 2N4, Canada

  • Venue:
  • Fuzzy Sets and Systems
  • Year:
  • 2008

Quantified Score

Hi-index 0.20

Visualization

Abstract

Many classification problems involve features whose specificity demand some form of feature space transformation (preprocessing) coupled with post-processing consensus analysis in order to accomplish a successful discrimination between different classes. In this study, we present a new methodology, which systematically addresses these design classification issues. At the preprocessing phase we offer a new approach of stochastic feature selection. This type of feature selection, collates quadratically transformed feature subsets for presentation to a collection of respective classifiers. In the sequel, independent classification outcomes are aggregated through fuzzy integration. The motivation behind the proposed methodology is twofold. Often, only a subset of features possesses discriminatory power while the remainder has a tendency to confound the effectiveness of the underlying classifier. Quite commonly, classification based on some consensus of classification outcomes coming from a set of classifiers operating upon different feature subsets becomes more accurate than the classification results produced by any individual classifier. To illustrate this design methodology, we discuss a classification problem coming from software engineering. Here we are concerned with a dataset comprosed of features describing a collection of qualitative attributes of a software system. The experiments demonstrate that the aggregated classification results using fuzzy integration are superior to the predictions from the respective best single classifiers.