Real-Time Detection of Credit Card Fraud in Imbalanced Datasets Using Machine Learning and Ensemble Approaches

The increase in digital payments has become more prominent and challenging in this short period of time and poses a great threat to both banks and customers. The traditional rule-based fraud detection systems cannot handle the evolving attack patterns, large volumes of data, and significantly biased transaction datasets. This work proposed a comprehensive machine learning-based fraud detection system that integrates advanced data preparation, efficient resampling approaches, and a wide variety of supervised learning algorithms to overcome these challenges. It compares seven machine learning models using Random Under-Sampling and Synthetic Minority Over-Sampling Technique. All the models went through the same preprocessing pipeline, which was used to enhance the reliability of the model through feature scaling and anomaly control. This paper used the Latency-Aware Accuracy Index, a metric that considers latency and outperforms the traditional metrics. It provides empirical evidence that boosting-based models such as LightGBM, CatBoost and XGBoost using SMOTE outperform classical classifiers with respect to prediction performance and inference time efficiency. It illustrates that accuracy alone is not sufficient for practical implementation; latency plays an equally important role in minimizing financial loss by ensuring a smooth customer experience. The results indicate that ensemble learning models much exceed the performance of standard linear classifiers in balanced datasets. XGBoost demonstrated superior performance with SMOTE oversampling, getting an accuracy of 0.9898, an F1-score of 0.9898, and a LAAI of 0.9889. This shows that it is strong across different balancing approaches. On the other hand, Logistic Regression and Naïve Bayes always got accuracy values around 0.50, which shows that they could not find the complex non-linear patterns that are typical of fraud detection.