FRAUD DETECTION IN BANKING DATA BY MACHINE LEARNING TECHNIQUES

Abstract

As technology developed and e-commerce services expanded, credit cards became one of the most popular payment
methods, resulting in a rise in the number of banking transactions. In addition, the significant rise in fraud
requires high banking transaction costs. As a result, detecting fraudulent activities has become a fascinating topic.
In this study, it examines the use of class weight-tuning hyper parameters to control the weight of legitimate and
fraudulent transactions. Specifically, it uses Bayesian optimization to optimize the hyper parameters while
preserving practical issues such as unbalanced data. It proposes weight-tuning as a pre-process for unbalanced
data, as well as Cat Boost and XG Boost to enhance the efficiency of the LightGBM method by taking into
account for the voting mechanism. To enhance performance even further, it applies deep learning to fine-tune the
hyper parameters, particularly proposed weight-tuning technique. It conducts experiments using real-world data to
test the proposed methods. In addition to the standard ROC-AUC, it utilizes recall-precision metrics to better cover
unbalanced datasets. Cat Boost, LightGBM, and XGBoost, logistic regression is evaluated individually using a 5-
fold cross- validation method. In addition, the majority voting ensemble learning technique is used to evaluate
the performance of the combined algorithms. The results show that the proposed methods outperform the
cutting-edge methods and achieve a significant improvement in performance.