Visualizing High Dimensional Classifier Performance Data

Abstract

Classifier performance evaluation, which typically yields a vast number of results, may be approached as a problem of analyzing high dimensional data. Conducting an exploratory analysis of visual representations of this evaluation data enables us to exploit the advantages of the powerful human visual capabilities. This allows us to gain insight into the performance data, interact with it and draw meaningful conclusions about the classifiers and domains under study. We illustrate how visual techniques, based on a projection from a high dimensional space to a lower dimensional one, enable such an exploratory process. Moreover, this approach can be viewed as a generalization of conventional evaluation procedures based on point metrics that necessarily imply a higher loss of information. Finally, we show that within this framework, the user is able to study the evaluation data from a classifier point of view and from a domain point of view, which is infeasible with traditional evaluation methods.