Credit Card Fraud Detection: A Realistic Modeling and a Novel Learning Strategy

Abstract

Detecting frauds in credit card transactions is perhaps one of the best testbeds for computational intelligence algorithms. In fact, this problem involves a number of relevant challenges, namely: concept drift (customers' habits evolve and fraudsters change their strategies over time), class imbalance (genuine transactions far outnumber frauds), and verification latency (only a small set of transactions are timely checked by investigators). However, the vast majority of learning algorithms that have been proposed for fraud detection rely on assumptions that hardly hold in a real-world fraud-detection system (FDS). This lack of realism concerns two main aspects: 1) the way and timing with which supervised information is provided and 2) the measures used to assess fraud-detection performance. This paper has three major contributions. First, we propose, with the help of our industrial partner, a formalization of the fraud-detection problem that realistically describes the operating conditions of FDSs that everyday analyze massive streams of credit card transactions. We also illustrate the most appropriate performance measures to be used for fraud-detection purposes. Second, we design and assess a novel learning strategy that effectively addresses class imbalance, concept drift, and verification latency. Third, in our experiments, we demonstrate the impact of class unbalance and concept drift in a real-world data stream containing more than 75 million transactions, authorized over a time window of three years.

Streaming active learning strategies for real-life credit card fraud detection: assessment and visualization

“…Two sets of experiments are performed: the first measures the detection accuracy at the level of the transactions, while the second measures the detection accuracy at the card level. In the first study, for the sake of simplicity, the classification model C is a conventional random forest model RF while a more realistic model A (discussed in [33] and in Section 3) is used for the cards 3 . Since the randomization process in RF and A may induce variability in the accuracy assessment, we present the results of twenty repetitions of the streaming.…”

“…For instance, Wei et al [51] introduce ContrastMiner, a fraud detection framework which combine the use of contrast pattern mining, neural network and decision forest to have a high precision in the detection. In our previous research we have used and assessed several binary classifiers for fraud detection [33,13]. Also one-class classifiers, like one-class SVM and Isolation Forest, belong to this category.…”

“…In the first experiment, the classifier C is implemented by a conventional Random Forest, while in the second, we use a more complex approach (ensemble of classifiers) dictated by the more challenging nature of the detection tasks. This approach has been presented in [5,33] and consists of the weighted average of two classifiers…”

“…This Feedback component is very important to address concept drift. This architecture is the result of an extensive model selection and assessment procedure which has been discussed in our previous work [5,33]. The aim of this paper is to discuss the impact of different AL strategies, so we will not take into consideration alternative classifier architectures.…”

Outlier Mining Methods Based on Graph Structure Analysis

“…This publicly available database [56][57][58][59][60][61] contains credit card transactions made in September 2013 by European cardholders. It contains 284807 transactions made in 2 days, of which 492 correspond to frauds.…”