Fraud Detection in Online Credit Card Transactions Using Deep Learning

This study focused on developing a more efficient hybrid fraud detection model than the hybrid model of the study it was based on using a combination of supervised and unsupervised algorithms, respectively Light Gradient Boosting Machine (LGBM) and Kernel Principal Component Analysis (KPCA), to tackle the temporal aspect of the data and better detect fraudulent activity in online credit card transactions. With the IEEE-CIS credit card fraud dataset, the researchers trained their model after utilizing preprocessing techniques such as categorical correlation, frequency encoders, and scalers to fully utilize the dataset in batches. The researchers made two versions of some models for thorough comparison and evaluation: Version 1 (V1) as a baseline model with standard parameters, and Version 2 (V2) with more updated data processing techniques and experimental parameters. Computing the Receiver Operating Characteristic-Area Under Curve (ROC-AUC) score in each model as the performance metric, the results show that compared to the ENS-XGB-LGBM model of the basis study with an ROC-AUC score of 0.951, the researchers have achieved their highest score of 0.95376 with their V2 IKPCA-LGBM model showing a noticeably improved fraud detection efficiency. The outcome shows that the V2 IKPCA-LGBM model with its dimensionality reduction and boosting algorithms surpasses the performance of the model of the basis study, however, the researchers recommended utilizing a balanced dataset, that further super parameter and hyperparameter tuning be done, as well as more time in testing the two versions of the models for them to be fully explored and improved upon.