On the Application of ROC Analysis to Predict Classification Performance Under Varying Class Distributions
Machine Learning, 58 (1):
25-32 (2005)Abstract
We counsel caution in the application of ROC analysis for prediction of classifier accuracy under varying class distributions. The heart of our contention is that in real-world applications variations of class distribution are likely to result from forces that affect the distribution of the attribute-values, rather than forces that directly affect the class distribution. In statistical terms, it is usually the class, rather than the attributes, that is the dependent variable. If the class distribution alters as an indirect consequence of changes in the distribution of the attribute values, rather than vice versa, performance estimates derived through ROC analysis may be grossly inaccurate.