Finding the Best Classification Threshold in Imbalanced Classification

Classification with imbalanced class distributions is a major problem in machine learning. Researchers have given considerable attention to the applications in many real-world scenarios. Although several works have utilized the area under the receiver operating characteristic (ROC) curve to select potentially optimal classifiers in imbalanced classifications, limited studies have been devoted to finding the classification threshold for testing or unknown datasets. In general, the classification threshold is simply set to 0.5, which is usually unsuitable for an imbalanced classification. In this study, we analyze the drawbacks of using ROC as the sole measure of imbalance in data classification problems. In addition, a novel framework for finding the best classification threshold is proposed. Experiments with SCOP v.1.53 data reveal that, with the default threshold set to 0.5, our proposed framework demonstrated a 20.63% improvement in terms of F-score compared with that of more commonly used methods. The findings suggest that the proposed framework is both effective and efficient. A web server and software tools are available via http://datamining.xmu.edu.cn/prht/ or http://prht.sinaapp.com/.