Machine Learning Applied to the INSS Benefit Request

The materialization of the universalization of social protection, foreseen in the Constitution of Brazil in the chapter of Social Security, with the tripod of Health, Welfare and Social Assistance, specifically in the scope of Welfare, occurs through the granting and maintenance of benefits to all Brazilians who need this protection, which generates a huge demand for millions of requests for annual benefits to the INSS, which is the operator of these services. Receiving and analyzing benefit requests, among other processes, in a timely manner and with assertiveness, is complex and challenging, whether due to the volume of millions of requests for annual benefits or the diversity of benefits available, the different grant criteria and the urgency that the nature of these requests requires for the maintenance of the applicants' lives. Within this context, this study aims to develop some models, using machine learning techniques, and select the best one, which can predict whether a certain benefit request will be granted or dismissed, offering an opportunity that can be used as yet another tool to help in the analysis of new benefit requests, opening space so that the dynamics of the analysis process can be directed in a more agile and assertive way. The data source for the construction of the models, in this work, was obtained from the INSS Open Data Portal, which are contained in the INSS Open Data Plan, with the monthly files of Defined Benefits (Granted or Unferred) in the period of December 2018 to June 2020. As scope of analysis, algorithms such as KNN, SVC, Decision Trees, Logistic Regression etc. were addressed. Models were also built using the techniques of Ensemble Bagging and Boosting, reaching a set of seventeen analyzed algorithms. The algorithm that achieved the best performance, using the F1 metric as a determinant, was the eXtreme Gradient Boosting Classifier (XGB) with 80%. With this, the model performs the prediction with approximately 84% Precision, 76% Sensitivity and 81% for AUC. As a result of the study, a model was obtained capable of making the prediction if a specific benefit requirement would be granted or rejected, based on the requirement data, with a performance within the expectations established in the objectives.