Graphical Features Selection Method

Abstract

The performance of a classification process depends heavily on the feature used in it. The traditional features/variables selection schemes are mostly developed from the model fitting point of view, which may not be good or efficient for classification purpose. Here we propose a graphical selection method, which allows us to integrate the information in the test data set, and it is suitable for selection useful features from high dimensional data set. We applied it to the Thrombin data set, which was used in KDD CUP 2001. By using the selected features from our graphical method and a SVM classifier, we obtained the higher classification accuracy than the results reported in KDD Cup 2001.