Feature selection based on star coordinates plots associated with Eigenvalue problems

Abstract

Feature selection consists of choosing a smaller number of variables to work with when analyzing high-dimensional data sets. Recently, several visualization tools, techniques, and feature relevance measures have been developed in order to help users carry out the feature selection. Some of these approaches are based on radial axes methods, where analysts perform backward feature elimination by discarding features that have a low impact on the visualizations. Similarly, in this paper, we propose a new feature relevance measure for star coordinates plots associated with the class of linear dimensionality reduction mappings defined through the solutions of eigenvalue problems, such as linear discriminant analysis or principal component analysis. We show that the approach leads to enhanced feature subsets for class separation or variance maximization in the plots for numerous data sets of the UCI repository. Lastly, in practice, the tool allows analysts to decide which features to discard by examining their relevance and by taking into account previous domain knowledge.