Comparison of Supervised and Unsupervised Fraud Detection

Abstract

Fraud Detection is a challenging problems in Machine Learning. The most commonly used evaluation metric for fraud detection which is a binary classification Machine Learning problem is Area Under the Receiver Operating Characteristic curve (AUROC). In this paper we will show that AUROC is not always the correct metric to evaluate the performance of a classifier with high imbalance. We shall rather show that Area Under the Precision Recall curve (AUPR) is a better evaluation metric for the same. We will compare and contrast various supervised as well as unsupervised approaches to optimize the Area under PR curve for fraud detection problem.