dc.description.abstract | This thesis addresses the problem of split agent deposit fraud in mobile money transactions and proposes a machine learning-based approach for its detection. Mobile money services have become increasingly popular in developing countries, where they provide convenient and accessible financial services. However, fraud in mobile money transactions poses a substantial threat to both customers and service providers. Split agent deposit fraud is a type of fraud in which a mobile money agent splits a deposit into multiple smaller amounts to evade detection. Detecting such fraudulent activities is crucial for maintaining service integrity and protecting customers. To tackle this problem, the study utilized a synthetic mobile money transaction dataset generated using the MoMTSim financial simulation platform. The study aims to identify the properties of mobile money transactions and split-agent deposit fraud, and to design, evaluate, and validate an ML model to detect such fraud. The data set is cleaned, resulting in a reduced data set of 1,048,575 transactions. The study evaluates three machine learning classifiers, namely Gradient Boosting Classifier, Random Forest, and Logistic Regression, based on various performance metrics such as accuracy, Matthew Correlation Coefficient (MCC), precision, recall, and F1-score. The results show that the Gradient Boosting Classifier outperforms the other two models, achieving high accuracy (0.99), MCC (0.97) and F1-score (0.98) for imbalanced mobile money data. However, for balanced mobile money data, we see an equal performance between gradient boosting and random forest classifiers. The proposed solution of using machine learning classifiers demonstrates promise in detecting split agent deposit fraud in mobile money transactions. These techniques can aid mobile money service providers in improving fraud detection, safeguarding customers, and maintaining the integrity of their services. | en_US |