O uso de técnicas de aprendizado de máquina em grafos para predição da popularidade de atos normativos brasileiros sobre segurança alimentar

Abstract

The growing demand and availability of large-scale data on Brazilian normative acts present challenging opportunities for society, particularly in the construction of computational systems that can learn, reason, and make inferences based on prior knowledge. In this context, knowledge bases are extremely important assets for the representation and automated reasoning of knowledge in various application domains. A relevant example is the inference of information from its representation in networks (knowledge graphs), which has gained academic and industrial notoriety in recent years. Therefore, this dissertation aims to develop an Artificial Intelligence model using Machine Learning techniques in graphs to predict the popularity of Brazilian normative acts specific to the area of food security. The development of the project required the implementation of a structured and comprehensive processing pipeline to perform detailed analyses and produce relevant results. The database used was composed of Brazilian normative acts focused on food security, obtained from the official Ministério da Agricultura e Pecuária (MAPA) website. The collection of normative data was carried out using web scraping techniques, which enabled the structured and systematic capture of 320 normative acts in PDF format. The study stages included: data processing, topic modeling, creation of a similarity matrix, construction of graphs and their characteristics, generation of graph embeddings, and experiments with Machine Learning models using embeddings and graph features, in addition to the application of the SHAP method. The results show that the analysis of normative acts through BERTopic allowed the identification of relevant topics, while the construction of graphs and the application of Machine Learning techniques enabled the prediction of the popularity of these normative acts. It is concluded that the applied methodology not only provides a detailed and robust analysis of the popularity of normative acts but also significantly contributes to the research field by demonstrating the effectiveness of Machine Learning techniques in graphs in the legal and normative context.