Automatic Unsupervised Feature Selection using Gravitational Search Algorithm

Abstract

Feature selection is an optimization problem that selects the features that have minimum redundancy and maximum relevance to improve the efficiency of algorithms. In this paper, we have proposed a novel automatic unsupervised feature selection method based on gravitational search algorithm, called AFSGSA (automatic feature selection using gravitational search algorithm). In contrast to most of existing unsupervised feature selection techniques, the proposed AFSGSA requires no prior knowledge of the data to be classified and number of features to be selected. AFSGSA determines the optimal number of features of the data-set on the run. Statistical property of data-set is used to develop a novel threshold setting concept to refine the features. A novel fitness function is also proposed to make the search more efficient. The performance of AFSGSA has been compared with recently developed well-known feature selection techniques. The experimental results reveal the efficiency and efficacy of the proposed feature selection technique over other existing techniques.