Impact of Box-Cox Transformation on Machine-Learning Algorithms

Abstract

This paper studied the effects of applying the Box-Cox transformation for classification tasks. Different optimization strategies were evaluated, and the results were promising on four synthetic datasets and two real-world datasets. A consistent improvement in accuracy was demonstrated using a grid exploration with cross-validation. In conclusion, applying the Box-Cox transformation could drastically improve the performance by up to a 12% accuracy increase. Moreover, the Box-Cox parameter choice was dependent on the data and the used classifier.