Impact of Deep Feature Synthesis on Deep Learning in Electronic Transaction Fraud Detection

Electronic transaction fraud has been a severe threat in recent years, causing substantial financial losses and devaluing the reputation of financial institutions. Various machine learning and deep learning models have been cited in the literature to detect electronic transaction fraud effectively. Feature selection is crucial in enhancing the performance of a machine learning model. Automatic feature engineering techniques outweigh the manual feature selection process regarding the time spent and adapting to the changing nature of the dataset. However, to the best of our knowledge, the impact of state-of-the-art feature selection techniques, such as Deep Feature Synthesis (DFS), on the performance of deep learning-based electronic transaction fraud detection has not been thoroughly studied. Our study uses the DFS algorithm to generate features automatically from the labelled credit card transaction dataset collected from open-source sites. The data is fed into a baseline deep learning model, i.e., Convolutional Neural Networks (CNN), in two separate cases- with DFS and without DFS. A comparison of results from both cases shows that CNN with DFS outperforms the standalone CNN significantly in terms of accuracy, precision, recall, and F1 scores.