Algoritmos de aprendizado de máquina na predição e avaliação de evasão de clientes em ambiente de produção

• Ano de defesa: 2021
• Autor(a) principal: Oliveira, Breno
• Orientador(a): Soares, Anderson da Silva
• Banca de defesa: Soares, Anderson da Silva, Soares, Telma Woerle de Lima, Sousa, Rafael Teixeira
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Ciência da Computação (INF)
• Departamento: Instituto de Informática - INF (RG)
• País: Brasil
• Palavras-chave em Português: Desvio de conceito; Auto machine laerning; Dados em stream
• Palavras-chave em Inglês: Machine learning; Algorithms in predicting; Evaluating customer evasion in a production environment
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/11522

The development of machine learning solutions involves several well-established stages. However, scientific studies have a concentration on stages such as data engineering, model training, and performance evaluation metrics. The advent of machine learning solutions implementation in business environments at an unprecedented level inspires the revisiting of some problems previously mentioned in the literature, but little explored. Among them, monitoring and evaluating the deterioration of the solution over time. During machine learning models training, it is assumed that the data not seen by the model in production presents the same distribution as the data used during the training stage. However, production models can decrease/lose performance as data changes over time. This phenomenon is defined in the literature as concept deviation. In this context, this work proposes a methodology that uses Auto Machine Learning with data stream learning capable of mitigating eventual concept deviations that may arise in the models implemented in a production environment. Real data from a customer avoidance problem (Churn) of a large-circulation regional newspaper were used. Three machine learning models were implemented using two methodologies: the proposed methodology called autoML-DS and the reference methodology that makes use of conventional model retraining. The results showed that the reference methodology presents performance losses of the implemented models, while the autoML-DS has its predictive capacity preserved. AutoML-DS was able to adapt the models over time, without having to perform a complete retraining, keeping small variations in the error rate.