Análise posicional de jogadores brasileiros de futebol utilizando dados GPS

• Ano de defesa: 2018
• Autor(a) principal: Gasparini, Randal
• Orientador(a): Álvaro, Alexandre
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus Sorocaba
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC-So
• Palavras-chave em Português: Aprendizado de Máquina; Futebol; Classificação; GPS; Classification
• Palavras-chave em Inglês: Machine Learning; Soccer
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::SISTEMAS DE COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/ufscar/9748

The professional soccer is always changing and is constantly searching tools and data to help the decision-making, providing tatics and techniques to the team. In Brazil, this sport goes to same way and the investiments are considerables. The One Sports is a company that capture GPS data from professional soccer players of some brazilian teams. This set of data has a lot of features and the One Sports asked if was possible to predict the ideal position of a player. Then, was firmed a cooperation between a academic study and a comercial company. This work find to understand a propose methods and techniques to predict the ideal position of soccer player, using machine learning algorithms. The database has more of one million of tuples. It was submited to pre-processing step, what is fundamental, because generated new features, removed incomplete and noisy data, generated new balaced dataset and delete outliers, preparing the data to execution of the algorithms k-NN, decision trees, logistic regression, SVM and neural networks. With the purpose to understand the performance and accuracy, some scenarios was tested. There was poor results when executed multi-class problems. The best results come from binary problems. The models k-NN and SVM, specifically to this study, had the best accuracy. It is important to note that SVM spent more than six hours to finish your execution, and k-NN used less than one and half minute to end.