A Machine Learning Framework for Mobile Money Fraud Detection Using Gaussian Naïve Bayes: A Transaction-Level Evaluation on the Synthetic Paysim Dataset

Source abstract

In Mobile money services are increasingly targeted by fraud, including SIM-related abuse such as SIM swapping, which undermines user trust, exposes sensitive financial data, and costs telecom operators and financial providers billions of dollars annually. Traditional rule-based detection systems have proven insufficient for identifying evolving fraud patterns, motivating the adoption of machine learning approaches. This study presents the design and implementation of a transaction-level mobile-money fraud screening system built around the Gaussian Naïve Bayes (GNB) classifier and evaluated on the synthetic PaySim dataset. The system was implemented following the Rapid Application Development (RAD) methodology using object-oriented design techniques. The dataset was split 80%/20% into training and held-out test sets; normalization was fitted on the training data, and the Synthetic Minority Oversampling Technique (SMOTE, sampling ratio 0.1) was applied to the training set only, leaving the test set at its natural class distribution. Furthermore, predicted fraud probabilities were converted to class labels using a decision threshold of 0.3 instead of the conventional 0.5. The developed model achieved 99.91% accuracy, 89.22% precision, 88.67% recall, 88.94% F1-score, 99.95% specificity, an AUC-ROC of 0.9425, and a false positive rate of 0.05%. The findings indicate that Gaussian Naïve Bayes provides an interpretable and computationally practical baseline for transaction-level fraud screening. However, becuase PaySim contains no SIM-specific signals such as SIM-swap events, device/IMSI changes, or OTP request patterns, the developed model should be understood as a general mobile-money transaction fraud screen rather than a dedicated SIM-fraud detector; the synthetic nature of the dataset and the conditional-independence assumption of Naïve Bayes further limit direct generalization to live telecommunications environments. Future work should incorporate real-world telecom and mobile-money data with genuine SIM-level features, adaptive fraud behaviour, deep learning or sequential models, and real-time stream processing.