Detecting Fraud in Public Procurement: A GMM-Based Approach to Analyzing Tender Data

Corruption and bid rigging in public procurement distort competition and increase the costs of products and services for public institutions, causing problems in different societal domains. The current availability of public data in digital format brings opportunities for applying machine learning to build solutions that help to deal with corruption. However, there are many challenges, like data sparsity and modeling complexity. Furthermore, confirmed cases of fraudulent tenders are limited, making applying traditional supervised learning techniques unfeasible. This work proposes a novel methodology for analyzing patterns using Gaussian Mixture Models (GMM) to identify suspicious bidding patterns when only a few fraudulent cases are known. Our methodology tests the similarity of unlabeled tenders, which can be fraudulent or not, with the fraudulent cases in different subspaces for defining a risk indicator. We run experiments in a dataset with tender data for acquiring heavy equipment purchases in which only a few cases are known as fraudulent. Results showed that our GMM-based methodology effectively provides a risk indicator ranking, highlighting risky tenders, making it a valuable tool for public agencies to enhance transparency and accountability in procurement.