Classifier Evaluation with Missing Negative Class Labels

摘要: The concept of a negative class does not apply to many problems for which classification is increasingly utilized. In this study we investigate the reliability evaluation metrics when contains an unknown proportion mislabeled positive instances. We examine how can inform us about potential systematic biases in data. provide motivating case and general framework approaching show that behavior unstable presence uncertainty labels stability depends on kind bias Finally, type amount present data have significant effect ranking degree they over- or underestimate true performance classifiers.