Investigating Unsupervised Machine Learning in Public Procurement Fraud Detection: A Case Study of POTRAZ Using the Isolation Forest Algorithm

Public procurement fraud poses a significant risk to public institutions, particularly in emerging economies where oversight mechanisms are often reactive and under-resourced. This study investigates the application of unsupervised machine learning—specifically, the Isolation Forest algorithm—to detect anomalies in procurement data that may indicate fraudulent activities. Using a dataset of 730 transactions from the Postal and Telecommunications Regulatory Authority of Zimbabwe (POTRAZ), the model was implemented, evaluated, and validated using expert-annotated labels and interpretability tools. The Isolation Forest achieved a precision of 94.6% and an F1-score of 69.3%, revealing its effectiveness in highlighting irregular transactions, although it exhibited limitations in recall (54.7%) for more subtle, context-driven anomalies. Visualisation via t-SNE and explainability through SHAP affirmed the model’s alignment with expert judgment. The study recommends a hybrid approach combining machine learning outputs with expert insights and proposes ensemble models for enhanced anomaly detection. These findings contribute to the growing body of work promoting AI-driven governance and fraud risk management in developing countries.