I know you're a fraud: Uncovering illicit activity in a Greek bank transactions with unsupervised learning

In recent years, the banking industry has undergone a significant transformation driven by the digitization of payments, which has led to a rise in financial and credit card fraud. Financial institutions face tremendous losses yearly in trying to mitigate this threat. Machine and deep learning techniques have become the spotlight in this ongoing battle, emerging as a very promising solution that many financial institutions are now actively adopting. Supervised methods (such as SVM, random forest, Markov models, etc.) have been widely studied in credit card fraud detection scenarios. However, their effectiveness is hindered by the need for labeled instances, which are both expensive and difficult to come by, and their lower performance when attempting to detect unknown attack patterns. On the other hand, unsupervised techniques do not require labeled instances and have proven to be effective in discovering new and evolving fraud patterns. This research paper employs a confidential industrial dataset composed of transactions made by cardholders of a Greek bank. We utilize two machine learning models, Isolation Forest, One-class SVM, and a deep neural autoencoder, to detect unsupervised anomalies in credit card data. This work introduces a novel approach to performing anomaly detection since our two machine learning models are trained on a per-cardholder basis, where a separate model is trained for each cardholder in our dataset. Moreover, the autoencoder is employed to identify frauds across the entire dataset, effectively acting as a generalized anomaly detection model. Finally, we examine the use of genetic algorithms for feature selection and additionally incorporate Shapley Additive Explanations (SHAP) to promote interpretability and transparency. Experimental results demonstrate the superiority of the autoencoder in comparison to the machine learning models in this anomaly detection scenario and stress the benefits of applying explainability tools such as SHAP to industrial models.