Research Specifications

Losses related to fraudulent transactions are increasing, so building a fraud detection system is essential. Previous studies have employed a variety of data mining and machine learning techniques to construct fraud detection systems. This study presents a new hybrid method based on the supervised autoencoder and the extreme gradient boosting (XGBoost) method. This combined method uses the power of a supervised autoencoder to generate an expressive representation of the data. It employs the XGBoost method as a robust classifier to detect fraudulent transactions. The hyperparameters of the proposed method are fine-tuned using the Bayesian optimization algorithm. The experiments on a public dataset containing 280 thousand records demonstrated that the proposed method achieves better results than the baseline method considering all the performance criteria, including Recall, Precision, and F1 measure.