Particle ranking: An Efficient Method for Multi-Objective Particle Swarm Optimization Feature Selection

This paper presents a novel multi-objective particle swarm optimization feature selection method. For this task, feature vectors are decoded as particles and ranked in a two-dimensional optimization space. To rank a particle, optimization space is modeled as two bands of dominated and non dominated particles with respect to that particle. Furthermore, uniform and nonuniform distributions of particles in optimization space as well as main properties of the proposed model are analyzed mathematically and experimentally in details. Beside particle ranks, feature ranks are also used to update velocity and position of particles in each iteration of optimization process.The proposed method has been evaluated in 16 datasets and compared with 11 state of the art feature selection and multi-objective optimization methods. Visual experimental results show that the proposed method finds Pareto Fronts of the best particles close to origin in multi-objective optimization space. Quantitative experiments also show that the proposed method achieves: the best Success Counting Measure in 13 datasets, superior C-Metric in 14 datasets, the greatest Hyper-Volume Indicator in 13 datasets. Finally, results of pairwise Mann–Whitney U-test show that the proposed method is statistically better in 38 pairwise statistically tests out of 55 tests.