Information-Theory-based Nondominated Sorting Ant Colony Optimization for Multiobjective Feature Selection in...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Feature selection (FS) is crucial for improving classification performance by identifying optimal feature subsets.
• Multiobjective optimization in FS aims to minimize features and maximize accuracy, facing challenges from feature interactions and discontinuous Pareto fronts.
• Ant Colony Optimization (ACO) shows promise for FS but lacks effective multiobjective approaches for complex datasets.

Purpose of the Study:

• To develop an effective Ant Colony Optimization (ACO)-based approach for multiobjective feature selection (FS).
• To address challenges in FS, including feature interactions and discontinuous Pareto fronts, using an information-theory-based method.
• To enhance classification performance and feature reduction through a novel ACO strategy.

Main Methods:

• An Information-theory-based Nondominated Sorting ACO (INSA) was developed for multiobjective FS.
• Modified ACO probabilistic functions using information theory to assess feature importance.
• Introduced a new ACO strategy for solution construction and a novel pheromone update mechanism for solution diversity.

Main Results:

• INSA demonstrated superior performance compared to machine learning, single-objective, and state-of-the-art multiobjective algorithms.
• Evaluated on 13 benchmark classification datasets with varying dimensionality.
• Achieved comparable or better classification performance with similar or fewer features than peer methods.

Conclusions:

• INSA effectively handles feature interactions and discontinuous Pareto fronts in multiobjective FS.
• The proposed information-theory-based approach enhances feature selection accuracy and efficiency.
• INSA offers a robust solution for optimizing feature subsets in classification tasks.