Credit Card Transactions, Fraud Detection, and Machine Learning: Modelling Time with LSTM Recurrent Neural Networks

In recent years, topics such as fraud detection and fraud prevention have received a lot of attention on the research front, in particular from payment card issuers. The reason for this increase in research activity can be attributed to the huge annual financial losses incurred by card issuers due to fraudulent use of their card products. A successful strategy for dealing with fraud can quite literally mean millions of dollars in savings per year on operational costs. Artificial neural networks have come to the front as an at least partially successful method for fraud detection. The success of neural networks in this field is, however, limited by their underlying design a feedforward neural network is simply a static mapping of input vectors to output vectors, and as such is incapable of adapting to changing shopping profiles of legitimate card holders. Thus, fraud detection systems in use today are plagued by misclassifications and their usefulness is hampered by high false positive rates. We address this problem by proposing the use of a dynamic machine learning method in an attempt to model the time series inherent in sequences of same card transactions. We believe that, instead of looking at individual transactions, it makes more sense to look at sequences of transactions as a whole; a technique that can model time in this context will be more robust to minor shifts in legitimate shopping behaviour. In order to form a clear basis for comparison, we did some investigative research on feature selection, preprocessing, and on the selection of performance measures; the latter will facilitate comparison of results obtained by applying machine learning methods to the biased data sets largely associated with fraud detection. We ran experiments on real world credit card transactional data using two innovative machine learning techniques: the support vector machine (SVM) and the long short-term memory recurrent neural network (LSTM). M. Bianchini et al. (Eds.): Innovations in Neural Infor. Paradigms & Appli., SCI 247, pp. 231–268. springerlink.com c © Springer-Verlag Berlin Heidelberg 2009 232 B. Wiese and C. Omlin 1 Fraud and Fraud Detection “Card companies continue to increase the effectiveness and sophistication of customerprofiling neural network systems that can identify at a very early stage unusual spending patterns and potentially fraudulent transactions [6]”. There are several different factors that make payment card fraud research worthwhile. The most obvious advantage of having a proper fraud detection system in place is the restriction and control of potential monetary loss due to fraudulent activity. Annually, card issuers suffer huge financial losses due to card fraud and, consequently, large sums of money can be saved if successful and effective fraud detection techniques are applied. Ultimately, card fraud detection deals with customer behaviour profiling in the sense that each card holder exhibits an ever evolving shopping pattern; it is up to the fraud detection system to detect evident deviations from these patterns. Fraud detection is therefore a dynamic pattern recognition problem as opposed to an ordinary static binary classification problem. The question here is, given a sequence of transactions, can a classifier be used to model the time series inherent in the sequence to such an extent that deviations in card holder shopping behaviour can be detected regardless of the skewness and noise inherent in the data ? In addition our classifier must exhibit both a high probability of detection and a low false alarm rate during generalisation; otherwise, it will be practically useless. Since we are ultimately dealing with a dynamic problem here, the question also arises whether a dynamic neural network machine learning technique will outperform a static one. This study aims at answering these questions by providing a comparative analysis on the use of support vector machines and long short-term memory neural networks for payment card fraud detection.