The Effect of Attribute Scaling on the Performance of Support Vector Machines

Abstract

This paper presents some empirical results showing that simple attribute scaling in the data preprocessing stage can improve the performance of linear binary classifiers. In particular, a class specific scaling method that utilises information about the class distribution of the training sample can significantly improve classification accuracy. This form of scaling can boost the performance of a simple centroid classifier to similar levels of accuracy as the more complex, and computationally expensive, support vector machine and regression classifiers. Further, when SVMs are used, scaled data produces better results, for smaller amounts of training data, and with smaller regularisation constant values, than unscaled data.