Anomaly Detection in Credit Card Transactions using Autoencoders

Abstract: Money lost in credit card fraud reached approximately 27.85 billion dollars worldwide in 2018. Using machine learning and anomaly detection, fraud detection can be utilised with the goal of solving this major problem. This thesis investigates whether Autoencoders can be used for fraud detection in credit card transaction and if they outperform Random Forest models in terms of AUROC score. Three different models were created: Random Forest, vanilla Autoencoder and LSTM Autoencoder. All models were trained on two different datasets, where the first consisted of real-life data and the second of synthetic data. The LSTM Autoencoder was trained in two different ways on the second dataset. One where data was sorted by time and one where data was sorted by users. A third model was then created combining the two LSTM Autoencoder models. All models were evaluated using accuracy, recall and AUROC. AUROC was the primary metric. The Random Forest model outperform the Autoencoder models on both datasets in terms of AUROC. The AUROC scores were fairly similar on the real-life dataset for all models, with the Random Forest model having the highest AU- ROC score of 0.9258. For the synthethic dataset the Random Forest model got an AUROC score of 0.8508 whilst the Autoencoder models got much lower AUROC scores between 0.6447 and 0.7921. The Autoencoders created in this thesis can be used for anomaly detection in credit card transaction data, but does not necessarily perform well, depending on the data used.