Investigation of the Noise Sensitivity of Machine Learning Algorithms on Credit Card Fraud Detection

Abstract

The misleading of machine learning based credit card fraud detection systems, due to various cyber attacks and information transfer-related distortions, is highly critical for the financial sector and its effects globally. In this study, the noise sensitivity and reliability of the machine learning algorithms on the credit card transactions database, which was balanced by over sampling method, were investigated. For this purpose, the noise generated in different distributions was added from 5% to 100 percent level and applied on different algorithms. Common noise distributions such as Normal, Poisson, Pareto, Exponential, Power and Uniform have been used. Logistic regression, K nearest neighbor, Decision trees, Random Forest, Extreme Gradient Boosting (XGB) and Gradient Boosting (GB) machine learning algorithms have been used in this study. Results were evaluated by complexity matrix and f1 score. The results include evaluation and comparison of classification criteria for each algorithm and noise level for the noise sensitivity study.