Credit card fraud detection remains a pressing challenge due to the rarity and complexity of fraudulent transactions. This paper presents a comparative review of twelve peer-reviewed studies that utilize the IEEE-CIS Fraud Detection dataset, a high-dimensional and imbalanced real-world benchmark. The reviewed methods span traditional machine learning, ensemble techniques, deep learning architectures, hybrid models, and privacy-preserving frameworks. Performance metrics such as AUC and F1-score are analyzed, with reported AUC values ranging from 0.88 to 0.9999. Key methodological trends include the dominance of ensemble methods, the emergence of unsupervised anomaly detection, and growing interest in federated learning. Despite high performance, issues like overfitting, lack of interpretability, and reproducibility persist. The paper concludes by identifying critical research gaps and proposing future directions in explainable AI, temporal modeling, and privacy-aware learning.