Utilização de R-GCN para identificação de conluio em auditorias de licitações públicas

Abstract

This work employs graph neural networks to identify anomalies in public procurement processes conducted by municipalities. The proposal is based on constructing a heterogeneous graph that integrates information about bidders, tenders, municipalities, and business partners, encompassing multiple types of relationships among these entities. Procurement processes are labeled according to a criterion that considers as suspicious of collusion those in which there is a pair of bidders with a high frequency of prior joint participations. The R-GCN (Relational Graph Convolutional Network) model was used to perform binary node classification in heterogeneous graphs under a supervised learning approach. It was observed that the approach with an undirected heterogeneous network achieved superior results. The results obtained reached an accuracy of approximately 96%.