Particle distance rank feature selection by particle swarm optimization

This paper presents a feature selection method in multi-objective particle swarm optimization space. For this task, a novel particle ranking is proposed based on particle distance from dominated and non-dominated particles and then used for feature rank computation. Position and velocity of particles are updated by a new update rule relies in feature ranks encoded in a vector. Properties of the proposed method are proven mathematically and supported in experiments. The proposed feature selection method is evaluated on 12 UCI datasets and 4 datasets from real-world applications compared with 5 state-of-the-art feature selection methods. As a visual comparison, the proposed method finds better non-dominated particles in two-dimensional optimization space with lower run time. Experiments also showed that the proposed method outperforms existing feature selection methods with regard to Success Counting Measure, C_Metric, Hyper-Volume Indicator and Statistical Analysis.

Automated summary

This paper introduces a feature selection method based on particle distance and feature ranking, which is mathematically proven and experimentally supported to outperform existing methods in multiple evaluation metrics.