Improved Competitive Swarm Optimization Algorithms for Feature Selection

CSO is an optimization algorithm based on the competition concept, which has been applied to feature selection and had a good performance on datasets with high feature dimensions. As a wrapper method, CSO is confronted with the problem of computation expensive and time consuming. To solve the problems, we proposed two improved algorithms from the perspective of reducing particle fitness calculation times: FCSO (Faster Competitive Swarm Optimization) and SFCSO (Selected Faster Competitive Swarm Optimization). FCSO shortens the time required for running by reducing the number of particles involved in each iteration, whereas SFCSO achieves the same goal by screening mechanism. SFCSO improves the stability of FCSO. We use KNN classifier to carry out experiments on four datasets with different size and dimension. The experimental results show that FCSO reduced time to one-tenth of the original, while SFCSO was half of it. FCSO has a better performance on binary classification problems so as SFCSO in multi-classification problems. Both algorithms can significantly reduce time complexity with a little decline of accuracy or even a higher accuracy, which is acceptable.