A new machine learning framework could help financial institutions identify suspicious credit card transactions with greater accuracy and resilience, according to a new study published in the same journal. Frontiers of artificial intelligence.
This research “Applying supervised machine learning algorithms and ensemble models to enhance credit card fraud detection.” We tested supervised ML models, resampling techniques, behavioral features, and ensemble learning techniques to improve fraud detection across a highly imbalanced credit card transaction dataset.
Fraud detection faces increasing pressure from digital payments and imbalanced data
The increased use of credit cards has expanded the attack surface for fraud, with research noting that global payment card fraud losses reached $27.85 billion in 2018 and are projected to rise to $35.67 billion by 2023. In the United States, reported fraud losses exceeded $10 billion in 2023, an increase of 14% from the previous year.
Credit card fraud is a threat not only to individual cardholders, but also to banks, merchants, governments, and payment networks. Fraud can include card theft, skimming, phishing, and fraudulent use of card details. As mobile banking, e-commerce, and digital wallets grow, fraud patterns are becoming faster and more complex, making it difficult for traditional rules-based systems to catch them.
Machine learning is increasingly being used to address this problem because it can learn patterns from past transaction data and classify new transactions as legitimate or suspicious. However, this study highlighted a major hurdle: fraud datasets are highly unbalanced. In a typical payment system, the number of legitimate transactions significantly exceeds the number of fraudulent transactions. In the primary dataset the researchers used, fraudulent transactions accounted for only 0.5% of records. This imbalance can distort model performance.
Although the system appears to be highly accurate by only classifying most transactions as legitimate, the most important and rare cases of fraud are still missed. For banks and payment processors, missed fraud can lead to financial losses, while excessive false alarms can disrupt genuine transactions and lead to customer dissatisfaction. Therefore, this study highlights the need to balance recalls that accurately capture fraud incidents with reducing unnecessary alerts.
To address this challenge, the researchers tested several supervised machine learning models: Decision Trees, Logistic Regression, Naive Bayes, Random Forests, Artificial Neural Networks, and XGBoost. We also applied three resampling techniques: random under-sampling, random over-sampling, and synthetic minority over-sampling techniques. These methods were used to reduce bias against legitimate transactions and improve the ability to detect rare fraud events.
This study used six datasets that combined synthetic data with real-world credit card and payment fraud data. The main training dataset contained 1.3 million synthetic transactions, and five unproven datasets were used to test whether the best model could generalize beyond the data used for training. Validation of this multiple datasets was the focus of the study, as many fraud detection studies rely on a single dataset, which may not indicate how the model performs under different data conditions.
Behavioral features and ensemble models improve detection performance
The researchers built a framework around a cross-industry standard process for data mining, a structured lifecycle for machine learning projects. This process includes understanding the business, understanding the data, preparing the data, modeling, optimizing, evaluating, and testing the invisible data.
Data preparation played a big role. In this study, feature transformation, encoding, scaling, data partitioning, and feature selection were applied. The researchers used both filter and wrapper techniques to identify the most relevant variables, including correlation-based selection, variance thresholding, ANOVA, Gini index, recursive feature removal, and forward feature selection. This hybrid approach aimed to remove irrelevant or redundant features while preserving variables that improve prediction.
Researchers created behavioral features designed to capture anomalous customer activity. These include transaction frequency, transaction timing, and anomaly score. The goal was to go beyond basic transaction attributes to detect deviations from the cardholder’s normal behavior. For example, a sudden spike in transactions or transaction activity that occurs outside of a user’s normal time frame can indicate potential fraud.
The results showed that the performance of the standalone models varied significantly. Logistic regression and Naive Bayes were weaker overall, while decision trees and artificial neural networks produced more acceptable results. Random Forest and XGBoost were among the most powerful standalone models, especially after resampling and thresholding.
Threshold optimization was used to improve the trade-off between precision and recall. The default classification threshold of 0.5 may not be optimal for fraud detection because identifying more fraud cases may be more important than maximizing standard accuracy. The researchers tested threshold changes for Random Forest and XGBoost and found that Random Forest using random oversampling with a 0.2 threshold and XGBoost using SMOTE with a 0.7 threshold produced a stronger fraud detection balance.
This study tested bagging, boosting, and stacking models. Ensemble learning combines multiple models to improve prediction stability and reduce reliance on a single algorithm. This is valuable in fraud detection, as the transaction patterns captured by the models may vary.
Among the ensemble approaches, the bagging model achieved the best overall performance. The selected Bagging 1 model combined decision tree, random forest, and artificial neural network learner under different resampling conditions. In the final summary of the study, we achieved precision of 0.99, recall of 0.90, and precision of 0.77, showing that we can identify most fraudulent transactions while keeping false positives to a manageable level.
Researchers found that bagging was more stable than boosting or stacking across key performance metrics. Boosting using Decision Tree and AdaBoost performed strongly, while other boosting configurations performed poorly. The stacking models also showed good performance, with one stacking model showing balanced precision and recall, and the other emphasizing recall. Nevertheless, the bagging model was chosen because it showed the most stable performance and better generalizability.
Adding behavioral features further improved the results. When new features were added to the training dataset, the Bagging 1 model’s F1 score increased from 0.79 to 0.83, precision increased from 0.73 to 0.77, and recall increased from 0.86 to 0.89. On the first unconfirmed dataset, the addition of the same features improved the F1 score and precision, and the recall remained stable at 0.88. These benefits suggest that behavioral patterns can enhance fraud detection beyond standard transactional variables.
Real-world deployments require monitoring, privacy protection, and reduction of false positives
This study argues that an integrated fraud detection framework can perform better than individual model improvements. Rather than focusing solely on algorithm selection, the researchers combined feature engineering, class imbalance handling, threshold tuning, ensemble diversity, and unseen data validation. This broader pipeline is positioned as a scalable approach for financial fraud detection.
The model showed good performance even on unknown datasets. On several unconfirmed datasets, the selected bagging model reached high precision, high recall and precision, and some datasets recorded scores above 0.95 across key metrics. This is important because real financial systems encounter different transaction patterns than the training data. A model that works well on only one dataset may fail when deployed to a live banking environment.
The researchers also acknowledged that there are practical limitations. Ensemble models can increase computational complexity, and real-time fraud detection requires low-latency systems that can quickly process large numbers of transactions. Bagging models are relatively scalable because they can be parallelized, but deployment may require model pruning, distributed computing, or incremental training.
False positives remain a major operational concern. Models with stronger recalls may catch more fraud, but if they flag too many legitimate transactions, banks can face customer dissatisfaction, transaction delays, and additional review costs. Therefore, this study recommends further work on thresholds and decision boundaries to balance fraud detection and operational costs.
Researchers also call for continuous monitoring of the model as fraud techniques evolve. Static models can lose effectiveness as criminals adapt to detection systems. Managing conceptual drift, where statistical patterns of transactions change over time, requires updating the actual financial system. This is especially important in digital banking, where fraud techniques can change rapidly across channels and geographies.
The study recommends more advanced feature engineering, stronger use of user behavior analysis, deeper investigation of bagging techniques, and scalable pipelines that can handle large numbers of transactions in real time. It also points to the need for collaborative research and responsible data sharing among researchers, banks, industry players, and regulators while protecting privacy and security.
Future research could also investigate the interpretability of the model. Financial institutions often need to explain why transactions are blocked or flagged, and regulators may require transparent decision-making in automated systems. Interpretability is especially important when machine learning models impact customer access to payments.
Researchers also propose extending fraud detection to network-based analysis to identify organized fraud groups. Social network analysis, which is already being studied in the context of anti-money laundering, could help detect relevant suspicious actors, rather than treating each transaction as a separate event.
