A hybrid machine learning model for detecting and preventing corruption in Kenya"s public procurement contracts.

Abstract

This study investigated the use of a hybrid machine learning model to detect and prevent corruption in Kenya’s public procurement contracts. Persistent procurement irregularities continue to undermine fiscal accountability, inflate contract prices and weaken governance despite having an established legal framework. To address these challenges, the study adopted a quantitative, experimental and data-driven research design, analyzing a substantially expanded dataset of 10,214 procurement transactions obtained from public data sources such as the Public Procurement Regulatory Authority Annual Procurement Reports, the National Treasury/IFMIS Open Contracting Portal, the Kenya Open Data Initiative, the Office of the Auditor-General audit reports and the EACC National Ethics and Corruption Survey datasets. The dataset incorporated structured fields (bid amounts, procurement method, number of bidders, supplier identifiers, award timelines) and textual indicators extracted from audit narratives and tender justifications. The study adopted a quantitative, data-driven and experimental research design integrating supervised and unsupervised machine learning. Logistic regression and random forest models were trained alongside K-Means clustering to detect hidden patterns of fraud. Data preprocessing using Python libraries involved cleaning, deduplication, normalization, missing-value imputation and natural language processing for extracting red-flag terms from unstructured reports. Model robustness was ensured through cross-validation and an 80/20 train-test split. Results showed that single-bid tenders, prior supplier allegations, persistent award extensions and contract value discrepancies were the strongest predictors of corruption likelihood. The random forest classifier achieved an AUC of 0.93, precision of 0.89 and recall of 0.86, while logistic regression recorded an AUC of 0.81 with strong interpretability value for policy justification. Unsupervised clustering successfully isolated high-risk contract groups hence validating the model’s anomaly detection capability in partially labelled environments. The findings indicated that integrating machine learning with cybersecurity principles such as data integrity checks, access control significantly enhanced the detection and prevention of corruption in procurement processes. The study concluded that corruption in public procurement was measurable and predictable using data- driven techniques. It recommended the integration of the hybrid model into Kenya’s Integrated Financial Management Information System and e-procurement platforms to facilitate real-time fraud detection and risk-based auditing. The study further recommended that procurement officers be trained on the use of predictive analytics and explainable artificial intelligence tools to interpret risk alerts effectively. Overall, the research contributed a practical, cybersecurity- enhanced analytical framework that strengthened transparency, accountability and governance in Kenya’s public resource management system.