Classificação de risco em transferências voluntárias federais utilizando XGBoost

• Autor(a) principal: GUILHON, Daniel Moreira
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: por
• Título da fonte: Biblioteca Digital de Teses e Dissertações da UFMA
• Texto Completo: https://tedebc.ufma.br/jspui/handle/tede/tede/3152
• topic: Transferências voluntárias; Aprendizagem computacional; XGBoost; Predição de risco; Voluntary transfers; Machine learning; XGBoost; Risk prediction; Ciência da Computação

After the Brazilian re-democratization, states and municipalities had to rely on federal government’s voluntary transfers of resources to achieve their public policies. For greater timeliness in the recovery of resources that may have been spent inappropriately, it is necessary to assign risk profiles of success or failure of these transfers. In this work, we propose a methodology that uses eXtreme Gradient Boosting (XGBoost) algorithm, using balanced and unbalanced data sets, with the use of hyperparameter optimization techniques, such as Tree-structured Parzen Bayesian Estimator (TPE). The results achieved good success rates. Results for XGBoost using balanced data showed a recall of 89.3% and unbalanced data a recall of 87.8%. However, for unbalanced data, the AUC score was 98.1%, against 97.9% for balanced data. Incorporating information data about the agreed object using natural language processing techniques can improve the results obtained.