Eficácia no uso de aprendizagem de máquina para estimação de risco em contratos públicos e empresas.

Abstract

The hiring of companies by the Public Administration is a frequent and necessary process in providing services to municipalities and government agencies in several countries. By law, contracts concluded must be accompanied by their execution. However, due to the large number of administrative contracts generated between companies and public entities, and the scarcity of management resources, it is impracticable to inspect all signed documents and their progress. One of the control strategies used by criminal prosecution and accountability agencies is the prioritization of cases to be investigated based on their level of risk. In the literature, the problem of risk estimation using machine learning for management entities is dealt with, in most cases, in the Federal context. Currently, in Paraíba, this choice of cases is made in practice through a series of heuristics that make up a Risk Matrix. Through experimental evaluation of the e ectiveness of this practical model and six machine learning models to the risk assessment of companies, we concluded that this approach had outperformed that one. In the context of public municipal contracts of the Paraíba state, through experimental analysis of the e ectiveness of three class balancing techniques and six models of machine learning, it was validated that the use of these models for public contracts is, indeed, promising. Also, we noted that building models with data from specific areas (in this work, engineering works and services) could improve its performance. With this, we aim to assist the use of public resources and the design of internal control strategies that allow subsidizing external control entities.