Signal processing on graphs for improving automatic credit card fraud detection

In this paper, several methods based on signal processing on graphs are proposed to improve the performance of credit card fraud detection. The proposed methods consist of a variant of the classic iterative amplitude adjusted Fourier transform (IAAFT) and two methods that we have called iterative surrogate signals on graph algorithms (ISSG). The objective is to generate surrogate samples from the original scarce fraud samples to improve the training of the detectors by lowering the variance of the estimate. A reliable augmentation of the target scarce population of frauds is important considering issues such as labeling cost; algorithm testing; data confidentiality; and constantly changing of patterns in the data streaming source. We have approached several scenarios with different legitimate and non-legitimate transaction ratios showing the feasibility of improving detection capabilities evaluated by means of receiver operating characteristic (ROC) curves and several key performance indicators (KPI) commonly used in financial business.