Hybrid Fraud Detection in Digital Transactions Using XGBoost and 1D Convolutional Neural Networks

Abstract

The increased popularity of digital payment systems has resulted in efficient and convenient online financial operations. On the other hand, such development requires more efficient fraud detection methods as more opportunities have appeared for fraudsters to conduct their illegal operations. This project suggests utilizing a hybrid method of fraud detection based on the use of PCA, XGBoost and CNN1D. PCA helps to remove unimportant features and improve the quality of input data. XGBoost algorithm is designed to detect complicated dependencies between the transaction features, while CNN1D is used to learn hidden features which can be used to distinguish real and fraudulent transactions. The experiment was conducted using the publicly available financial transaction dataset divided into train and test data samples in 70:30 proportion. Accuracy, precision, recall, F1-score and confusion matrix were used to evaluate the performance of the model. The obtained experimental results prove that the suggested hybrid model is capable of detecting fraud accurately and reliably.