Credit Card Fraud Detection using One-Class Classification Algorithms

Abstract

Similar to most things in everyday life, the advent of payment cards also has good and bad sides. It undoubtedly, made life easier by bringing the whole payment system to a single card, but it also paved the way for a new set of illegal activities and frauds. The credit card fraud has been carried out since the payment cards came into existence, and since then, the trend in such frauds has been an increasing one. Therefore, a quest to attenuate the losses caused by such frauds began. For this purpose, many preventive and detective measures have been taken in the past, and new ways are sought to further improve the policies. These measures, however, reduce the losses temporarily only and have not yet succeeded in converting the uptrend in the losses by such frauds into a downtrend because fraudsters always come up with a new way of tricking the people and the system. Thus, a new way of solving this ever-existing challenge is needed, which can detect even those fraudulent instances that are executed by techniques and methods that are yet-to-be-invented by fraudsters. Moreover, the occurrence of normal (non-fraudulent) credit card transactions is much more than fraudulent ones, and therefore, the data for credit card fraud detection is highly imbalanced. Another challenge in credit card fraud detection systems is the high dimensionality of datasets. Therefore, to address the imbalance nature of the data, to cope with the curse of dimensionality with a new way of making the model to regulate and extract the discriminative features, and to detect the fraud carried out by yet-to-be-invented techniques, we implemented a set of novel and state of the art subspace learning-based One-Class Classification algorithms. We experimented with integrating a projection matrix and geometric data information in the training phase to improve credit card fraud detection. We also experimented by using a maximization-update rule in updating the projection matrix instead of the classical minimization-update rule in the subspace leaning-based data description. We found that the linear version of Graph-embedded Subspace Support Vector Data Description with kNN graph, gradient-based solution, and minimization-update rule works better than all other models.