Generic performance measure for multiclass-classifiers

The evaluation of classification performance is crucial for algorithm and model selection. However, a performance measure for multiclass classification problems (i.e., more than two classes) has not yet been fully adopted in the pattern recognition and machine learning community. In this work, we introduce the multiclass performance score (MPS), a generic performance measure for multiclass problems. The MPS was designed to evaluate any multiclass classification algorithm for any arbitrary testing condition. This measure handles the case of unknown misclassification costs and imbalanced data, and provides confidence indicators of the performance estimation. We evaluated the MPS using real and synthetic data, and compared it against other frequently used performance measures. The results suggest that the proposed MPS allows capturing the performance of a classification with minimum influence from the training and testing conditions. This is demonstrated by its robustness towards imbalanced data and its sensitivity towards class separation in feature space.