Feature selection algorithm with feature space separability estimation using discriminant analysis

this paper seeks to address a range of feature selection methods used in intelligence analysis tasks. The methods usually help to find a set of features that optimize classification procedure, enhance machine learning algorithms accuracy or reduce dimensionality. Generally speaking, data is comprised of a vast number of features, a lot of which are irrelevant for the task of interest. Application of some specific features may improve both accuracy and performance of algorithms. Such features are usually named as relevant features. This paper proposes a new feature selection algorithm based on estimation of feature space separability. Separability criteria is calculated by using discriminant analysis. In the paper an opportunity to select relevant features faster than existing methods with the proposed algorithm has been empirically shown. Experiment shows that a set of relevant features collected by the means of the proposed algorithm, in contrast to previous feature selection algorithms, allows to build more accurate classification model.