Comparison the Quality of Classification Algorithms

This article summarises and describes different measures of the predictive quality (performance) of classification models. Quantitative measures of quality can be complemented with the visual representation. One of the best known is the ROC (Receiver Operating Characteristic) curve that represents the performance of binary classification model within the full range of conditions (costs and class distributions) of discrimination. The duality between ROC and expected costs of a classifier (cost curve) is described in more details here. Cost curve measures the difference in performance of two classifiers directly in expected costs (EC). The use of ROC and EC is highly important when asymmetric misclassification costs, imbalanced probabilities of classes or changing conditions occur. A hybrid classifier can be found on the convex hull (ROCCH) of the ROC curves of different classifiers. This is potentially the best classification model for any mixture of outside constraints. Lower envelope of cost curves of different classifiers corresponds to ROCCH in ROC space.