Performance Variability of Machine Learning Models using Limited Data for Collusion Detection: A Case Study of the Brazilian Car Wash Operation

Fraudulent companies form illegal agreements, like collusion and cartels, to circumvent the impartiality and competitiveness of the public procurement auctions. These types of fraud can cause significant financial losses and erode trust in the public sector. Therefore, building reliable methods for early detection of frauds is a priority for public organizations. This study uses an enriched version of the “Operation Car Wash” dataset to evaluate the collusion detection capabilities of different machine learning algorithms. Using cross-validation techniques, the methodology proposed in our work was able to improve the collusion detection rate of the learning models used in this work, outperforming the results of other works found in the literature.