Improving linear discriminant analysis with artificial immune system-based evolutionary algorithms

  • Authors:

  • Mehdi Mohammadi;Bijan Raahemi;Ahmad Akbari;Babak Nassersharif;Hossein Moeinzadeh

  • Affiliations:

  • Iran University of Science and Technology, Computer Engineering Department, University Road, Hengam Street, Resalat Square, Tehran, Iran;University of Ottawa, 55 Laurier Ave. E., Ottawa, ON, Canada K1N 6N5;Iran University of Science and Technology, Computer Engineering Department, University Road, Hengam Street, Resalat Square, Tehran, Iran;Electrical and Computer Engineering Department, K.N. Toosi University of Technology, Iran;Iran University of Science and Technology, Computer Engineering Department, University Road, Hengam Street, Resalat Square, Tehran, Iran

  • Venue:
  • Information Sciences: an International Journal
  • Year:
  • 2012

Quantified Score

Hi-index 0.07

Visualization

Abstract

Mapping techniques based on the linear discriminant analysis face challenges when the class distribution is not Gaussian. While using evolutionary algorithms may resolve some of the issues associated with non-Gaussian distribution, the solutions provided by evolutionary algorithms may get trapped in local optimum. In this paper, we propose a hybrid approach using evolutionary algorithms to improve the accuracy of linear discriminant analysis. We apply combinations of the artificial immune system and fuzzy-based fitness function to address the cases with non-Gaussian distribution classes, and at the same time, evade local optimum of the search space. The transformation matrix computed by fuzzy-based evolutionary algorithms is used during the preprocessing step of the classification process to map the original dataset into a new space. The proposed methods are evaluated on datasets selected from UCI, as well as a network dataset collected from real traffic on the Internet. We measure five different indexes, namely mutual information, Dunn, SD, isolation and DB indexes to evaluate the extent of the separation of the samples before and after the proposed mapping is performed. The mapped datasets are then fed to some different classifiers. Then, accuracy of the pre-processing methods are observed on different classifiers (with and without proposed mapping). The experimental results demonstrate that the fuzzy fitness-based evolutionary methods outperform other previously published techniques in terms of efficiency and accuracy.