A Model for Detecting Fraud in Banking Transactions Based on Artificial Intelligence: Gambling-Related Transactions

Abstract

Banking fraud is one of the major challenges that can have significant economic consequences for society. The aim of this study was to classify and detect gambling-related activities using banking transaction data. The dataset consisted of 16,764 banking transactions collected between 2023 and 2024, belonging to 1,857 distinct bank card numbers, with the deposit status column considered as the target label. Following data preprocessing procedures, including data cleaning, normalization, categorical-to-numerical conversion, and standard scaling, statistical features were extracted from two primary variables, resulting in the generation of 44 new features. Due to the high dimensionality of the feature space and the limited dataset size, three dimensionality reduction approaches based on statistical feature-selection tests were employed. The dataset was divided into training and testing subsets using an 80:20 ratio. In the baseline model, the K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Random Forest, Extreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Convolutional Neural Network (CNN) algorithms were evaluated using the original feature set. The proposed model was an ensemble learning framework that combined XGBoost and Random Forest classifiers through a soft-voting mechanism. Model hyperparameters were optimized using a greedy search strategy. The results demonstrated that the statistical test-based feature selection method achieved the best performance among the dimensionality reduction techniques, with an accuracy of 80.38%. Under this feature-selection approach, the proposed ensemble learning model also achieved an accuracy of 80.38%, representing an improvement of more than 6% compared with the best-performing baseline model (XGBoost with an accuracy of 73.92%). Furthermore, deep neural networks exhibited overfitting throughout all experimental stages, indicating that the available data volume was insufficient for deep learning approaches. Overall, targeted feature engineering, statistical test-based dimensionality reduction, and ensemble learning constitute an effective approach for classifying banking transactions when data availability is limited.