Reducing corruption in public procurement using machine learning

Public procurement processes are prone to corruption in many countries around the world. Many regulatory agencies that monitor and improve procurement use a manual and reactive review process. This makes it slow and expensive to detect and reduce corruption. In this work, we partnered with the National Agency for Public Procurement (DNCP) in Paraguay to take their current manual procurement review process and augment it with machine learning to complement the regulator’s work in automating reviews. Our system analyzes procurements when they are created and generates a risk score on whether procurement is likely to receive complaints (a proxy for corruption) in the future or not. Compared to a review on a first-come-first-served basis, we show that reviews based on our classification model improve the complaints capture rate from 30 percent to 78 percent. In addition, we consider equity, and balance efficiency with policy goals to reduce bias towards high value procurements in the review process. The result is a more efficient process, that is less bias towards high value procurements.