Use of Machine Learning for Active Public Debt Collection with Recommendation for the Method of Collection Via Protest

Abstract

This work consists of applying supervised Machine Learning techniques to identify which types of active debts are appropriate for the collection method called protest, one of the means of collection used by the Attorney General of the State of Pernambuco. For research, the following techniques were applied, Neural Network (NN), Logistic Regression (LR), and Support Vector Machine (SVM). The NN model obtained more satisfactory results among the other classification techniques, achieving better values in the following metrics: Accuracy (AC), FMeasure (F1), Precision (PR), and Recall (RC) with indexes above 97% in the evaluation with these metrics. The results showed that the construction of an Artificial Intelligence/Machine Learning model to choose which debts can succeed in the collection process via protest could bring benefits to the government of Pernambuco increasing its efficiency and effectiveness.