Machine Learning–Based Classification of Mobile Money Transactions

Abstract

Mobile money has become the backbone of everyday payments across much of the Horn of Africa, and the transaction records it leaves behind are a valuable but under-used source of insight. This paper investigates whether the type of a mobile money transaction can be recovered automatically from its attributes, a capability that would support transaction monitoring and customer analytics. Working within the Knowledge Discovery in Databases framework and using the WEKA platform, we compared three well-established classifiers—Naïve Bayes, J48 (WEKA's implementation of the C4.5 decision tree), and Random Forest—on a real dataset of 3,731 transactions described by ten attributes and labelled with one of ten transaction types. All models were assessed under 10-fold stratified cross-validation. The two tree-based methods clearly outperformed the probabilistic baseline: J48 reached 76.01% accuracy with a Kappa of 0.6147 and the lowest mean absolute error (0.0615), while Random Forest was almost identical at 75.58% and produced the smallest root mean squared error (0.1821); the difference between the two tree-based models was not statistically distinguishable. Naïve Bayes, hampered by its independence assumption, managed only 51.09%. A per-class breakdown showed near-perfect recognition of common categories such as P2P receive (F-measure above 0.93) but a complete failure on the rarest classes, several of which scored zero. This imbalance effect is quantified by the macro-averaged F1 of just 0.40 and balanced accuracy of 38.02% for the best model, far below its 76.01% overall accuracy. This gap is evidence that headline accuracy is inflated by the dominant classes. On balance we recommend J48 for deployment under the present experimental setting, since it pairs competitive accuracy with fast training and human-readable rules that translate directly into monitoring logic.