Aplicação de algoritmos não supervisionados em dados eleitorais

• Ano de defesa: 2019
• Autor(a) principal: Polizeli, Mateus Vendramini [UNIFESP]
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Paulo (UNIFESP)
• Palavras-chave em Português: Anomaly Detection; Outliers; Electoral Data; Isolation Forest; K-Means.; Detecção De Anomalias; Dados Eleitorais; K-Means
• Link de acesso: https://sucupira.capes.gov.br/sucupira/public/consultas/coleta/trabalhoConclusao/viewTrabalhoConclusao.jsf?popup=true&id_trabalho=7928459; https://repositorio.unifesp.br/handle/11600/59852

Given the incessant search of society for clarity in government spending, management efficiency and transparency using the public agency, the structuring of works that allow a thorough investigation to efficiently monitor these actions becomes relevant. From an initial study in the literature, it was verified the existence of a series of controls and disclosure of accountability of sectors and public agencies. However, despite initiatives such as these, there is still little work considering further investigation to capture possible irregularities in the policy instrument. Thus, the objective of this project is to study some mechanisms for detecting anomalies associated with the 2018 electoral candidate data set. The proposed methodologies are based on unsupervised algorithms K-Means and Isolation Forest in an attempt to create a decision support tool for regulators to direct human resources for research. A combination of these algorithms, referred to here as KM+IF, is also suggested in order to improve accuracy and decrease the error rates associated with the models. The results observed in this project indicate that the proposal KM+IF shows good performance for situations where the variables of interest are available. However, it may yield unsatisfactory results when they are not available. In the case study for the set of electoral candidates, the overall result of the KM+IF algorithm was lower than the individual result of the K-Means and Isolation Forest techniques.