Avaliação de algoritmos de aprendizado de máquina na classificação de lançamentos bancários: um estudo aplicado ao Ministério Público do Estado da Bahia

Abstract

This study investigated the effectiveness of several machine learning algorithms in the task of classifying bank transactions, using data received from financial institutions by the Public Prosecutor's Office of the State of Bahia (MPBA). The general objective of the study was to strengthen the MPBA's investigative capacity by automating and improving the process of classifying and analyzing financial transactions. The methodology employed focused on a comparative approach to the performance, in executing the proposed task, of traditional machine learning techniques: Logistic Regression, Support Vector Machine (SVM), Decision Tree and K-Nearest Neighbors (KNN) and also of various adjustments of the pre-trained BERT model. The data used in this study were anonymized in accordance with legal requirements and rigorously explored and dimensioned in order to meet the technical demands and the available computational capacity. The main results revealed that, in general, the models studied presented excellent performance in classifying bank transactions. The SVM model, with the appropriate configuration, achieved an F1-score of 88.54% on the test dataset, outperforming the other models tested. This study contributes significantly to the adoption of artificial intelligence solutions in the public sector, proposing the foundations of a scalable model that can be applied in other public institutions for the analysis of financial data. The application of these techniques not only improves existing operations, but also opens new avenues for innovation in the processing and analysis of large volumes of data in government agencies.