A new performance evaluation method for two-class imbalanced problems

Abstract

In this paper, we introduce a new approach to evaluate and visualize the classifier performance in two-class imbalanced domains. This method defines a two-dimensional space by combining the geometric mean of class accuracies and a new metric that gives an indication of how balanced they are. A given point in this space represents a certain trade-off between those two measures, which will be expressed as a trapezoidal function. Besides, this evaluation function has the interesting property that it allows to emphasize the correct predictions on the minority class, which is often considered as the most important class. Experiments demonstrate the consistency and validity of the evaluation method here proposed. © 2008 Springer Berlin Heidelberg.