Comparative Analysis of ANN and CNN Using SMOTE and SHAP for Fraud Detection
Year: 2026
Type: conference-paper
Abstract: Credit card fraud remains among the most serious security issues in the worldwide financial network, making it important to develop more effective fraud detection algorithms capable of dealing with complex data sets. The purpose of this research is to provide a comparison between Artificial Neural Network (ANN) and Convolutional Neural Network (CNN) approaches for detecting credit card fraud. A major obstacle in this study is the class imbalance where the legit transactions outnumber the fraudulent ones in the dataset. This research implemented a hybrid approach using Synthetic Minority Over-sampling Technique (SMOTE) to get rid of the class imbalance before training the models. Methods include pre-processing of the Credit Card Fraud Detection dataset on Kaggle, followed by evaluation of the two network architectures. This experiment confirmed that even though both architectures performed well, with high accuracy, the ANN architecture was shown to have greater robustness after hyperparameter tuning. After being tuned, the ANN architecture recorded a recall rate of 0.706, an F1-Score of 0.516, and a ROC-AUC of 0.977. On the other hand, the CNN architecture regressed in performance, showing reduced classification performance after tuning. In order to enhance interpretability of the results generated, the SHAP tool was used for interpretation of non-linear results. It is concluded that the SMOTE-ANN architecture can be said to be more sensitive.
Institution: Binus University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7214654249