Algorithm for Vascular Access in our Imaging Department

Abstract

Measuring the ability of detection models to predict abnormal or fraudulent events is an important feature of any framework for assessing financial crime, and the Gini coefficient is one of the most common metrics used for assessing the accuracy of a classification model. Traditionally used in inequality studies and credit risk modeling, the Gini coefficient has very powerful features that can be applied to assess the discriminative power of fraud-detection models. Compared to other applications financial crime data set are limited in terms of fraud, which creates many issues around the relevance of performance metrics. The computation of confidence intervals has become a crucial step especially for models built with limited data.Resampling approaches were the first techniques used to compute variance for the Gini coefficient. Some authors have shown that the estimation of Gini’s coefficient can be obtained from ordinary linear regression (OLS) based on the data and its ranks, thereby also providing an exact analytic standard error. These techniques can be developed for assessing the quality of fraud detection models and for measuring the confidence interval of the Gini coefficient. Special attention needs to be given to low fraud rate and/or small length data sets and low-quality models. A new sampling-based method (F-Gini) is proposed for measuring the standard error of the Gini coefficient more adapted to these situations.