Performance Evaluation of Machine LearningModels Via Mobile Payment for FraudIdentification

Abstract

Digital and mobile payments have not only led to a surge in fraudulent activities detected in financial systems but also made
detecting fraud more difficult. The conventional rule-based methodology frequently lacks the ability to rely on complex patterns of
fraud, and thus results in high FP and FN. Using the extremely skewed PaySim dataset, this research presents a machine learning
model for detecting mobile money transfer fraud and laundering. A trained XGBoost classifier was then used to learn complex
transactional relationships, and overfitting was checked by using regularization built in. The model was tested with accuracy (ACC),
precision (PRE), recall (REC) and F1-score (F1) and ROC-AUC metrics all reaching 99.6%, 99.8%, 98.7% and 0.991%, respectively.
The ROC curve and confusion matrix prove that there is a high ability to discriminate and low levels of false alarms. As compared
with the recent methods like Bert and DenseNet, it is evident that XGBoost can perform significantly better. These findings indicate
that the suggested XGBoost-based model is a highly scalable, dependable, and efficient model to detect real-time fraud in the mobile
payment system and help a financial institution decrease losses and improve its anti-money laundering compliance.