Precision–Recall Trade-offs in Imbalanced Credit Card Fraud Detection: A Threshold Sensitivity Study

Credit card fraud detection continues to be a complex challenge in data mining, primarily due to thehighly imbalanced nature of real-world transaction datasets, where fraudulent cases represent only avery small fraction of total transactions. In such scenarios, conventional classification models oftenreport high overall accuracy while failing to effectively detect fraudulent activities. As a result, accuracyalone is not a reliable indicator of model performance, and greater emphasis must be placed on metricssuch as precision and recall.This study investigates the effectiveness of various techniques for handling class imbalance in creditcard fraud detection, with a particular focus on understanding the trade-off between precision and recall.A comparative analysis is conducted using baseline models, including Logistic Regression and DecisionTree, along with data-level approaches such as Random Oversampling, Random Undersampling, andthe Synthetic Minority Oversampling Technique (SMOTE). In addition, an algorithm-level strategybased on class weighting is employed to enhance minority class detection without altering the datasetdistribution.To further analyse model behavior, threshold sensitivity analysis is performed by adjustingclassification thresholds and observing the resulting changes in precision and recall. Experimentsconducted on a highly imbalanced credit card dataset (approximately 0.17% fraudulent transactions)demonstrate that techniques such as SMOTE and class weighting significantly improve recall, reachingvalues as high as 0.95. However, this improvement comes at the cost of a notable decline in precisiondue to an increase in false positives. Threshold tuning further intensifies this effect, clearly illustratingthe inherent trade-off between fraud detection capability and false alarm rates.Overall, the results indicate that no single technique provides a perfect balance between precision andrecall. Instead, effective fraud detection requires selecting an appropriate trade-off based on the specificrequirements and constraints of real-world applications.