Particle swarm optimization for feature selection in classification: a multi-objective approach.

Area of Science:

• Machine Learning
• Data Mining
• Computational Intelligence

Background:

• Classification tasks often involve numerous features, many of which are irrelevant or redundant, potentially hindering performance.
• Traditional feature selection methods often overlook the inherent multi-objective nature of optimizing classification accuracy while minimizing feature count.

Purpose of the Study:

• To present the first investigation of multi-objective particle swarm optimization (PSO) for feature selection.
• To develop and evaluate two novel PSO-based algorithms for generating Pareto fronts of optimal feature subsets.

Main Methods:

• Developed two distinct multi-objective PSO algorithms for feature selection.
• Algorithm 1 incorporates nondominated sorting into PSO.
• Algorithm 2 integrates crowding, mutation, and dominance concepts into PSO.
• Compared algorithms against conventional, single-objective, two-stage, and established evolutionary multi-objective methods on 12 benchmark datasets.

Main Results:

• Both PSO-based algorithms successfully evolved sets of nondominated feature subsets.
• The first algorithm surpassed conventional, single-objective, and two-stage methods, performing comparably to existing multi-objective algorithms.
• The second algorithm demonstrated superior performance compared to the first algorithm and all other evaluated methods.

Conclusions:

• Multi-objective PSO is a viable approach for feature selection, effectively addressing the conflicting objectives of maximizing performance and minimizing feature dimensionality.
• The proposed algorithms offer improved feature selection capabilities, particularly the second algorithm which shows enhanced effectiveness.