Network-based high level classification: novel models and applications

• Ano de defesa: 2021
• Autor(a) principal: Colliri, Tiago Santos
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Automação de investimentos; Bills voting; CBC test; Classificação de dados de alto nível; Complex networks; Corruption prediction; COVID- 19; COVID-19; Detecção de insuficiência; Hemograma; High level data classification; Insufficiency detection; Machine learning; Mercado de ações; Partidos políticos; Political parties; Predição de corrupção; Redes complexas; Stock market; Stock trading automation; Votações legislativas
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-26032021-102400/

Machine learning is an application of artificial intelligence with focus on the development of computer programs that can access data and use them to learn for themselves. High level data classification is a technique based on data pattern formation, instead of only their physical features. Complex networks have been proven to be quite useful for characterizing relationships among data samples and, consequently, they are a powerful mechanism to capture data patterns. In this work, we investigate novel ways of using the network-based approach in the development of high level classification techniques. Initially, two classification techniques are introduced, and their performances are assessed by applying them to benchmark datasets, both artificial and real, as well as comparing their results to those achieved by traditional classification models, on the same data. Afterwards, we explore the inherent advantages offered by this type of approach, such as its versatility and interpretability, by developing novel network-based techniques specifically designed to be applied on data concerning real and relevant problems from very diverse fields, from the financial market to corruption among politicians and healthcare. Although these type of applications certainly require a greater amount of effort from the part of researchers, in terms of the challenge and data preprocessing, we believe they are important to bring academic research closer to the reality. Among our findings, there is the uncovering of an unexpected relationship between legislative voting data and convictions for corruption or other financial crimes among Brazilian representatives. We also demonstrate how one can adapt a model, which originally has been applied to detect periodicity in meteorological data, for identifying up and down trends in the stock market, automatically triggering a buying or a selling order for the asset, accordingly. In another investigation, a technique to help healthcare workers in the task of monitoring COVID-19 patients is presented, by detecting early signs of hepatic, renal or respiratory insufficiency solely based on Complete Blood Count (CBC) test results. In summary, we believe this work makes an important contribution to the advance of large scale public data study using complex networks.