New Feature Engineering Framework for Deep Learning in Financial Fraud Detection

The total losses through online banking in the United Kingdom have increased because fraudulent techniques have progressed and used advanced technology. Using the history transaction data is the limit for discovering various patterns of fraudsters. Autoencoder has a high possibility to discover fraudulent action without considering the unbalanced fraud class data. Although the autoencoder model uses only the majority class data, in our hypothesis, if the original data itself has various feature vectors related to transactions before inputting the data in autoencoder then the performance of the detection model is improved. A new feature engineering framework is built that can create and select effective features for deep learning in remote banking fraud detection. Based on our proposed framework 19, new features have been created using feature engineering methods that select effective features based on their importance. In the experiment, a real-life transaction dataset has been used which was provided by a private bank in Europe and built autoencoder models with three different types of datasets: With original data, with created features and with selected effective features. We also adjusted the threshold values (1 and 4) in the autoencoder and evaluated them with the different types of datasets. The result demonstrates that using the new framework the deep learning models with the selected features are significantly improved than the ones with original data.