Learning suitable data representation for credit card fraud detection algorithms

With the recent pandemic, credit card and even contactless payment have gained significant popularity. The elevated frequency of card usage, along with the lack of diligence among customers, has resulted in an increase in stolen or counterfeit cards, often leading to fraudulent activities. This emphasized the importance of real-time detection of abnormal banking transactions for card issuers. Automated analysis of transaction logs is a prevalent approach to address this challenge, often involving the comparison of incoming transactions to a database containing genuine and fraudulent transactions. Methods such as the Mahalanobis distance have proven to be efficient when seeking for similarities between high dimensional data. However, the challenge lies in the fact that credit card logs contain both categorical and tabular data, which poses compatibility issues with the Mahalanobis algorithm. This study explores the effectiveness of finding alternative data representations as a pre-processing step to enable the utilization of algorithms that were previously unsuitable for such data types. The research conducted in this work is established on an exclusive credit card logs dataset.