Multiobjective GAs, quantitative indices, and pattern classification

Area of Science:

• Machine Learning
• Computational Intelligence
• Data Mining

Background:

• Single objective classifiers often struggle with overfitting and ignoring smaller classes.
• Approximating complex linear and nonlinear class boundaries requires robust methods.

Purpose of the Study:

• To develop a nonparametric genetic classifier using multiobjective optimization (MOO) to address limitations of single objective classifiers.
• To enhance classification by simultaneously minimizing misclassified points and hyperplanes while maximizing recognition scores.

Main Methods:

• Integration of MOO with variable length chromosomes for a genetic classifier.
• Introduction of validation sets and validation functionals for solution selection.
• Incorporation of elitism and domain-specific constraints, termed the CEMOGA-Classifier (constrained elitist multiobjective genetic algorithm based classifier).

Main Results:

• The CEMOGA-Classifier effectively approximates linear and nonlinear class boundaries.
• The classifier overcomes overfitting and the issue of ignoring smaller classes.
• New quantitative indices (purity, minimal spacing) were developed for evaluating MOO techniques.

Conclusions:

• The CEMOGA-Classifier offers a robust approach to classification, outperforming related methods.
• The developed MOO techniques and evaluation indices contribute to advancing classification algorithms.
• This method provides a significant improvement for handling complex datasets and imbalanced classes.