Machine Learning Applications In Fraud Detection For Digital Banking

Abstract

The rapid growth of digital banking has enhanced customer convenience while simultaneously exposing institutions to increasingly complex fraud schemes that evade traditional rule-based detection. To address this challenge, this study investigates the application of machine learning models to detect fraudulent transactions using the publicly available Financial Transaction for Fraud Detection Research dataset comprising 10,000 anonymized records. A structured methodology was adopted, beginning with data preprocessing, exploratory analysis, and temporal feature engineering, followed by the implementation of logistic regression, random forests, and gradient boosting algorithms. Class imbalance, a central challenge in fraud detection, was mitigated using cost-sensitive weighting and Synthetic Minority Oversampling Technique (SMOTE). Performance evaluation employed precision, recall, F1-score, ROC–AUC, and average precision metrics to capture model effectiveness under imbalanced conditions. Results demonstrated that while simple indicators such as transaction amount or device type were insufficient on their own, machine learning models achieved strong discriminative power, with ROC–AUC of 0.89 and average precision of 0.81. Precision (0.71) and recall (0.64) highlighted a practical yet improvable trade-off, underscoring the importance of threshold tuning and recall-focused strategies. The study concludes that machine learning offers not only technical viability but also operational necessity for fraud detection in digital banking, while emphasizing the need for adaptive retraining and explainability to sustain long-term effectiveness.