Multiobjective Differential Evolution for Feature Selection in Classification.

Area of Science:

• Computational Intelligence
• Machine Learning
• Data Science

Background:

• Feature selection is crucial for reducing dimensionality and improving classification accuracy in real-world problems.
• Existing methods often fail to identify multiple optimal feature subsets, despite their potential for similar performance.
• The non-uniqueness of optimal feature subsets necessitates methods capable of exploring diverse solutions.

Purpose of the Study:

• To develop a multiobjective differential evolution approach for discovering multiple optimal feature subsets.
• To enhance the diversity and quality of feature selection solutions.
• To address the limitation of existing methods that focus on a single optimal subset.

Main Methods:

• A multiobjective differential evolution algorithm is proposed for feature selection.
• An initialization method incorporating feature relevance is introduced for better starting points.
• A clustering technique divides the population into subpopulations, with subarchives using a crowding distance for diversity.
• An external archive retains nondominated solutions guided by an improved hypervolume contribution indicator.

Main Results:

• The proposed approach successfully evolves a superior Pareto front of feature subsets.
• Experiments on 14 datasets demonstrate improved performance compared to seven state-of-the-art methods.
• The method effectively finds diverse feature subsets yielding similar or identical classification performance.

Conclusions:

• The developed multiobjective approach effectively addresses the non-uniqueness of optimal feature subsets.
• It offers a significant advancement in evolutionary feature selection by preserving solution diversity.
• The method provides a more comprehensive understanding of the feature selection landscape for classification problems.