Best feature selection using successive elimination of poor performers
This paper addresses the issue of feature extraction and selection, focusing particularly ion the feature selection issue. Without assuming any particular classification algorithm it suggests that first one should extract as much information (features) as conveniently possible and then apply the proposed successive elimination process to remove redundant and poor features and then select a significantly smaller, yet useful, feature subset that enhances the performance of the classifier. The algorithm is formally described and is successfully applied to a four class ECG classification problem. A minimum distance classifier (MDC) using Mahalanobis distance as the decision criterion is developed. Using MDC an overall recognition performance of 87.5% is obtained on the testing set of the four ECG classes.
