Cost-sensitive thresholding over a two-dimensional decision region for fraud detection

Abstract

Credit fraud poses a challenging task in terms of detection. It can result in significant losses depending on the amount, so a cost-sensitive perspective needs to be taken. Classical approaches focus on estimating the probability of fraud and selecting a decision threshold, but they often fail to consider the transaction amount or account for the cumulative losses incurred within the sample. Consequently, these approaches can result in sub-optimal strategies. A new thresholding approach is proposed, based on the construction of a two-dimensional decision space with an estimated probability and the credit amount. This expansion allows more freedom for the optimal classification rule search, which is performed with a new algorithm. The proposed method generalizes previous approaches, so an improvement is consistently achieved. In addition, it allows a restricted search. This is shown in a study of two real data sets, comparing the results obtained by a wide range of classifiers.