Skip to the main content

Zbornik Istarskog veleučilišta - Rivista dell’ Università Istriana di scienze applicate, Vol. 2. No. 1., 2023.

Professional paper

Credit Card Fraud Detection Using Machine Learning Algorithms

Ivan Lorencin ; Istrian University of Applied Sciences
Nikola Anđelić ; University of Rijeka
Deni Vale ; Istrian University of Applied Sciences
Marko Mavrinac ; Electroindustrial and trade school Rijeka

Full text: english pdf 1.056 Kb

page 138-161

downloads: 221

cite

Abstract

One of the main challenges to the security of an online business is credit card fraud. For this reason, algorithms based on artificial intelligence and machine learning are being introduced to enable the most accurate and fast detection of card fraud. This paper presents an approach to the detection of card fraud based on machine learning algorithms more specifically, a multilayer perceptron (MLP) and a Decision tree. The aforementioned algorithms were trained and tested using a publicly available
data set on card fraud. The data set used consists of 7 characteristics of the card transaction and information on whether there was card fraud or not. In total, the data set contains information on 1,000,000 transactions, and it is highly imbalanced. To handle the class imbalance, random undersampling, SMOTE, and SMOTE-Tomek algorithms were proposed. From the achieved results it can be seen that the highest performances are achieved if MLP (AUC = 0.99, f1 = 0.99, MCC = 0.98, and Kappa = 0.98) and Decision tree (AUC = 0.99, f1 = 0.99, MCC = 0.99, and Kappa = 0.98) are trained by using data set re-sampled by using SMOTE-Tomek algorithm. If the performance of the mentioned algorithms is examined using fewer characteristics of the transaction, it can be seen that by reducing the number of characteristics a significant decrease in classification performances can be noticed if a Decision tree in combination with SMOTE-Tomek is used. However, if an MLP in combination with SMOTE-Tomek is used, a significantly lower decrease in performance can be observed, pointing to the higher robustness to input vector dimension reduction. Such a robust system can provide information about transaction validity even in a condition where the input data is limited to a few input variables. From the achieved results, it can be concluded that MLP in combination with the SMOTE-Tomek algorithm can be used for credit card fraud detection, even in conditions with a lower number of input variables.

Keywords

Hrčak ID:

309176

URI

https://hrcak.srce.hr/309176

Publication date:

24.10.2023.

Visits: 527 *