Multi-label feature selection based on competitive swarm optimization

Feature selection is one of the important preprocessing steps in data mining and machine learning, which is used to dimensionality reduction and selecting a subset of representative features. Feature selection can remove redundant and irrelevant features that can increase the accuracy of the machine learning tasks. In this paper, a novel embedded approach for the multi-label feature selection method using competitive Swarm Optimizer (CSO) is proposed. In this method, at first, of particles is generated, then the particles are divided into two equal groups and compete in pairs, the winners are moved to the next iteration and the losers learn from the winners, and at the end of each iteration, the objective function for all the particles is computed. To increase the convergence rate, half of the initial population is generated by the similarity between features and labels, and a local search method inspired by the gradient descent algorithm is applied to discover the local structure of data. Finally, based on the best particle, the feature selection is done. The performance of the proposed method is compared with six known and state-of-the-art multi-label feature selection methods. The experimental results on the image and text multi-label datasets show the efficiency and superiority of the proposed method in different multi-label evaluation measures criteria.