AI Spots Credit Card Fraud with 98.3 Per Cent Accuracy

Credit card fraud detection frequently suffers from imbalanced datasets, hindering the reliability of predictive models used in financial systems. Reza E. Fazel from EN Bank, Arash Bakhtiary from ReDi School, and Siavash A. Bigdeli from DTU, along with their colleagues, present a novel workflow to address this issue by optimising Explainable Boosting Machines (EBMs). Their research significantly advances the field by achieving a superior balance between predictive accuracy and model interpretability, allowing for not only precise fraud detection but also valuable insights into the factors driving those predictions. Through systematic hyperparameter tuning, feature selection, and a refined preprocessing pipeline utilising the Taguchi method, the team demonstrated a peak ROC-AUC of 0.983, exceeding previous EBM benchmarks and consistently outperforming established algorithms such as Logistic Regression, Random Forest, XGBoost, and Decision Tree models.

This breakthrough addresses a critical challenge in financial systems: accurately detecting fraudulent transactions within severely imbalanced datasets. Unlike conventional fraud detection methods reliant on potentially biased sampling techniques, this research introduces a workflow that prioritizes both predictive power and interpretability.

The optimized EBM effectively balances accuracy and transparency, enabling precise identification of fraudulent activity alongside actionable insights into the factors driving those predictions. Central to this advancement is the systematic refinement of the EBM through meticulous hyperparameter tuning, strategic feature selection, and optimized data preprocessing.
Researchers employed the Taguchi method to determine the optimal sequence for data scaling and model parameter settings, ensuring robust and reproducible performance gains. This work demonstrates the potential of interpretable machine learning to enhance trustworthiness in financial systems.

The EBM’s ability to rank feature importance and reveal interaction effects allows for a deeper understanding of fraud patterns, facilitating more effective risk management strategies. By avoiding data resampling or anomaly detection, the model maintains data integrity while delivering superior performance.

The findings suggest a pathway towards more reliable and transparent fraud detection, ultimately strengthening customer trust and the stability of financial institutions. The research focused on optimising EBM through systematic hyperparameter tuning, feature selection, and preprocessing refinement to address the inherent class imbalance in credit card data.

Rather than employing conventional sampling techniques, this approach aimed to achieve a balance between predictive accuracy and model interpretability. Data preprocessing involved a carefully designed sequence of scalers, optimised using the Taguchi method to maximise performance and reproducibility.

This method systematically varied the order of scaling techniques, evaluating their impact on the EBM model’s ability to discriminate between legitimate and fraudulent transactions. The research employed the Taguchi method to systematically refine both data scaling sequences and model hyperparameters, resulting in robust and reproducible performance gains.

Experimental evaluation utilized benchmark credit card data, revealing the optimized EBM consistently outperformed alternative machine learning models. This study focused on achieving high accuracy alongside interpretability, allowing for precise detection of fraudulent transactions and providing actionable insights into feature importance.

The Taguchi method ensured a systematic validation of performance improvements, optimizing the sequence of data scalers and model hyperparameters for maximum effectiveness. Systematic hyperparameter tuning, feature selection, and preprocessing refinement enhance the model’s performance while maintaining interpretability.

This approach achieves a balance between accuracy and the ability to understand which factors drive predictions, offering actionable insights into fraudulent transaction patterns. By identifying the 18 most influential features from an initial set of 30, the optimized EBM reduces computational cost and model complexity without sacrificing predictive power. The authors acknowledge that the combinatorial nature of hyperparameter tuning limited the scope of explored parameter configurations.

Future research could investigate alternative orthogonal array configurations to potentially achieve further performance gains. These findings establish a new benchmark in credit card fraud detection and highlight the potential of interpretable, data-driven optimization for trustworthy fraud analytics and broader applications in imbalanced classification problems.