Algoritmos de aprendizado de máquina aplicados na estimação de descontos em itens da cesta básica

Abstract

Purchases made by government entities in Brazil represent around 12% of the national GDP. Recently, the legislation governing these acquisitions was updated, emphasizing the importance of principles such as planning, effectiveness and economy. Machine learning (ML) techniques can be used by public managers in their efforts to comply with these new principles and trends in public administration in Brazil. In this regard, this study aimed to develop and evaluate machine learning models aiming to estimate the discount value offered in bids for items in the national basic food basket carried out by entities in the state of Paraná. Random Forest, XGBoost and Artificial Neural Networks techniques were used, evaluating their results using the mean absolute error (MAE) and root mean squared error (RMSE) metrics. The dataset was obtained through the “Portal de Informações para Todos” provided by the Paraná State Court of Auditors. Challenges such as the way data is made available and the lack of standardization have made preprocessing more complex, requiring the use of complementary techniques such as clustering. After filtering through the 18 basic food basket items analyzed and preprocessing the data, 18 attributes and 73,861 instances remained, which were used to train the models. Models were trained for each of the items, obtaining an average MAE of 0.099196 and an average RMSE of 0.122152, with the best result obtained using XGBoost, which achieved an RMSE of 0.049 for the "French Bread" item. ML models can provide relevant information for managers and regulatory agencies, assisting in budget planning, fraud detection and price negotiation in tenders, contributing to more economical and transparent public management.