A Hybrid COA-CNN Framework for Credit Card Fraud Detection Using Hyperparameter Optimization and Metaheuristic Feature Selection

Abstract

Credit card fraud identification remains a difficult problem due to the enormous class imbalance and the evolving nature of fraudulent behavior. The traditional machine learning methods, including Support Vector Machines (SVMs), and even a basic deep learning model, are often sensitive to the parameters of the employed algorithms, are constrained by scaling issues, and highly dependent on data resampling. To address these drawbacks and apply the optimization-based learning, the present paper will introduce a new hybrid framework which will compare the Cuckoo Optimization Algorithm (COA) with the Convolutional Neural Network (CNN). The COA is an optimizer that is two-way, and may be deployed either as automated hyperparameter optimization or as metaheuristic feature selection such that the model may be very high discriminative without necessarily aggressive oversampling or intricate network structural forms. The provided strategy was trained strictly on the Kaggle credit card fraud detection data set and Precision, Recall, F1-score, and ROC-AUC metrics were used to evaluate the performance. As the experiment results indicate, the state-of-the-art ROC-AUC of the COA- CNN is 0.990 and the model is more balanced in regards to the detection and computational performance. This information shows that nature-inspired metaheuristic optimization is efficient in enhancing the resilience and scalability of financial security systems based on deep learning.