The Difference Between Precision-recall and ROC Curves for Evaluating the Performance of Credit Card Fraud Detection Models

The study is devoted to the actual problem of fraudulent transactions detecting with use of machine learning. Presently the receiver Operator Characteristic (ROC) curves are commonly used to present results for binary decision problems in machine learning. However, for a skewed dataset ROC curves don’t reflect the difference between classifiers and depend on the largest value of precision or recall metrics. So the financial companies are interested in high values of both precision and recall. For solving this problem the precision-recall curves are described as an approach. Weighted logistic regression is used as an algorithmlevel technique and random undersampling is proposed as data-level technique to build credit card fraud classifier. To perform computations a logistic regression as a model for prediction of fraud and Python with sklearn, pandas and numpy libraries has been used. As a result of this research it is determined that precision-recall curves have more advantages than ROC curves in dealing with credit card fraud detection. The proposed method can be effectively used in the banking sector.