Abstract
Fraud detection systems support advanced detection techniques based on complex rules, statistical modelling and machine learning. However, alerts triggered by these systems still require expert judgement to either confirm a fraud case or discard a false positive. Reducing the number of false positives that fraud analysts investigate, by automating their detection with computer-assisted techniques, can lead to significant cost efficiencies. Alert reduction has been achieved with different techniques in related fields like intrusion detection. Furthermore, deep learning has been used to accomplish this task in other fields. In our paper, a set of deep neural networks have been tested to measure their ability to detect false positives, by processing alerts triggered by a fraud detection system. The performance achieved by each neural network setting is presented and discussed. The optimal setting allowed to capture 91.79% of total fraud cases with 35.16% less alerts. Obtained alert reduction rate would entail a significant reduction in cost of human labor, because alerts classified as false positives by the neural network wouldn't require human inspection.
Highlights
As new payment methods become widely available and credit card customer base grows, the volume of incoming transactions to be processed by an FDS increases, and so does the number of alerts to be reviewed by fraud analysts
In our research, deep neural networks were assessed from a perspective of credit card fraud alert reduction
A set of alerts triggered by an FDS were classified as either valid alerts, representing real fraud cases, or wrong alerts, representing false positives, by ten neural network architectures
Summary
As new payment methods become widely available and credit card customer base grows, the volume of incoming transactions to be processed by an FDS increases, and so does the number of alerts to be reviewed by fraud analysts. Several techniques have been designed to deliver decision support and false positive minimization in the intrusion detection field, which is very similar in nature to fraud detection. These techniques include adaptive learning [1], similarity with verified alerts [2], greedy aggregation algorithm [3], neuro-fuzzy approach [4], alert enrichment framework [5], and outlier detection [6]. ALAC (Adaptive Learner for Alert Classification) implemented adaptive learning through a classifier that captured decision patterns from analyst’s feedback. The obtained results showed that false positives were reduced by approximately 30% when tested against DARPA 1999 dataset [16]
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
