Improving classification accuracy through feature selection

Abstract

High accuracy is essential to any data mining process. A large part of the factors which influence the success of a data mining problem reside in the quality of the data used. Feature selection represents one of the tools which can refine a dataset before presenting it to a learning scheme. This paper analyzes a wrapper approach for feature selection, with the purpose of boosting the classification accuracy. A wrapper is viewed as a 3-tuple consisting of a generation procedure, an evaluation function and a validation procedure. Experimental evaluations have been performed for several combinations of the three components. The results have shown that feature selection improves the classification accuracy and speeds up the training process. Moreover, two robust combinations are proposed: one that constantly achieves highest accuracy, and one which significantly boosts the initial accuracy of the inducer.