Improving classification accuracy based on class-space reduction
Byungjoon Park, Sejong Oh · International Journal of Information and Communication Technology · 2015
Classification is a major topic in the field of information engineering and researchers have made many attempts to improve classification accuracy. Classification accuracy is highly dependent on overlapping areas between classes of a given dataset. In general, a larger overlap area produces lower classification accuracy. In this study, we suggest a new method to improve classification accuracy based on class-space reduction. Our proposed method rescales training/test data by moving data points in the direction of the centroid of the class to which the data points belong. By conducting experiments using real datasets, we confirm that the classification accuracy of many rescaled datasets generated by class-space reduction is improved for some classification algorithms.