Autonomous visualization

Abstract

Many classification algorithms suffer from a lack of human interpretability. Using such classifiers to solve real world problems often requires blind faith in the given model. In this paper we present a novel approach to classification that takes into account interpretability and visualization of the results. We attempt to efficiently discover the most relevant snapshot of the data, in the form of a two-dimensional scatter plot with easily understandable axes. We then use this plot as the basis for a classification algorithm. Furthermore, we investigate the trade-off between classification accuracy and interpretability by comparing the performance of our classifier on real data with that of several traditional classifiers. Upon evaluating our algorithm on a wide range of canonical data sets we find that, in most cases, it is possible to obtain additional interpretability with little or no loss in classification accuracy. © Springer-Verlag Berlin Heidelberg 2006.