Loading...
Publication
Técnicas de aprendizado de máquina aplicadas na previsão de desempenho de operadores de centros de teleatendimento
| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 1.81 MB | Adobe PDF |
Authors
Advisor(s)
Abstract(s)
Automatizar e melhorar a gestão e o desempenho profissional dos trabalhadores é um grande desafio enfrentado por várias empresas, sobretudo as que possuem um elevado número de funcionários. Portanto, as ferramentas de previsão de desempenho podem fornecer mecanismos especialmente importantes para o planejamento e gestão de recursos humanos das instituições intensamente dependentes do trabalho dos seus colaboradores.
Esta tese foca o uso de tecnologias de aprendizado de máquina integradas num pipeline dinâmico que contempla a manipulação e seleção dos dados de entrada e as parametrizações dos algoritmos utilizados para otimizar a previsão do desempenho dos trabalhadores no serviço de teleatendimento.
O trabalho de previsão de desempenho, definido neste trabalho pelo absenteísmo e produtividade, foi desenvolvido para uma população-alvo pertencente a uma grande empresa de prestação de serviços de teleatendimento brasileira. As variáveis foram extraídas do perfil dos agentes de teleatendimento e, em seguida, filtradas por processos de correlação e seleção de variáveis. Mais precisamente, neste trabalho, foram extraídas características pessoais, sociais e profissionais de teleatendentes para prever o desempenho de uma população de, aproximadamente, 10,5 mil funcionários.
Foram testados alguns modelos de previsão, para os quais um conjunto vasto de variáveis de entrada foram dinamicamente selecionadas e submetidas, permitindo assim comparar o desempenho obtido pelos vários algoritmos de aprendizado máquina utilizados (cf. LR - Logistic Regression, LSTM - Long Short Term Memory, MLP - Multilayer Perceptron, NB - Naive Bayes, RF - Random Forest, SVM - Suport Vector Machine e XGBoost - Extreme Gradient Boosting). A hiper-parametrização destes modelos de aprendizado de máquina também foi considerada na comparação dos algoritmos mais adequados para o problema de previsão. As hiperparametrizações, assim como a seleção de variáveis, foram ajustadas através do uso de um algoritmo evolutivo, permitindo melhorar os resultados de previsão que globalmente foram bastante promissores.
O conjunto das técnicas aplicadas no trabalho permitiu melhorar o entendimento sobre o problema da previsão de desempenho da empresa. Foi, assim, possível desenvolver um estudo intensivo de aplicação dos vários algoritmos de aprendizado máquina à previsão de desempenho. Este estudo foi suportado por mecanismos dinâmicos de seleção de variáveis de entrada mais significativas, bem como de técnicas de hiper-parametrização dos algoritmos, que melhor resultados de previsão produziam, de modo a selecionar os algoritmos de aprendizagem máquina mais adequados para o efeito.
Automating and improving the management and professional performance of workers is a major challenge faced by several companies, especially those with a high number of employees. Therefore, performance forecasting tools can provide important mechanisms for the planning and management of human resources of institutions that are intensely dependent on the work of their employees. This thesis focuses on the use of machine learning technologies integrated in a dynamic pipeline that contemplates the manipulation and selection of input features, and the parameterization of algorithms used to optimize the prediction of workers’ performance in the call center service. The forecasting of work performance, defined in this work by absenteeism and productivity, was developed for a target population belonging to a large Brazilian call center service company. The input features were extracted from the profile of the call center agents and then filtered through processes of correlation and selection of variables. More precisely, in this work, personal, social and professional characteristics of call center agents were extracted to predict the performance of a population of approximately 10,500 employees. Some forecasting models were tested, for which a vast set of input features was dynamically selected and submitted, allowing to compare the performance obtained by the various machine learning algorithms used (cf. LR - Logistic Regression, LSTM – Long Short Term Memory, MLP - Multilayer Perceptron, NB - Naive Bayes, RF - Random Forest, SVM - Suport Vector Machine e XGBoost - Extreme Gradient Boosting). The hyperparametrization of these machine learning models was also considered when comparing the most suitable algorithms for the forecasting problem. The hyper-parameterizations, as well as the selection of variables, were adjusted through the use of an evolutionary algorithm, allowing to improve the forecast results that were very promising. The set of techniques applied in this work improved the understanding about the forecasting problem for the company workers performance. It was possible to develop an intensive study through the application of various machine learning algorithms for performance prediction. This study was also supported by dynamic mechanisms for selecting the most significant input features, as well as the best algorithms’ hyper-parameterization, which produced better forecasting results, in order to select the most suited machine learning algorithms for the job.
Automating and improving the management and professional performance of workers is a major challenge faced by several companies, especially those with a high number of employees. Therefore, performance forecasting tools can provide important mechanisms for the planning and management of human resources of institutions that are intensely dependent on the work of their employees. This thesis focuses on the use of machine learning technologies integrated in a dynamic pipeline that contemplates the manipulation and selection of input features, and the parameterization of algorithms used to optimize the prediction of workers’ performance in the call center service. The forecasting of work performance, defined in this work by absenteeism and productivity, was developed for a target population belonging to a large Brazilian call center service company. The input features were extracted from the profile of the call center agents and then filtered through processes of correlation and selection of variables. More precisely, in this work, personal, social and professional characteristics of call center agents were extracted to predict the performance of a population of approximately 10,500 employees. Some forecasting models were tested, for which a vast set of input features was dynamically selected and submitted, allowing to compare the performance obtained by the various machine learning algorithms used (cf. LR - Logistic Regression, LSTM – Long Short Term Memory, MLP - Multilayer Perceptron, NB - Naive Bayes, RF - Random Forest, SVM - Suport Vector Machine e XGBoost - Extreme Gradient Boosting). The hyperparametrization of these machine learning models was also considered when comparing the most suitable algorithms for the forecasting problem. The hyper-parameterizations, as well as the selection of variables, were adjusted through the use of an evolutionary algorithm, allowing to improve the forecast results that were very promising. The set of techniques applied in this work improved the understanding about the forecasting problem for the company workers performance. It was possible to develop an intensive study through the application of various machine learning algorithms for performance prediction. This study was also supported by dynamic mechanisms for selecting the most significant input features, as well as the best algorithms’ hyper-parameterization, which produced better forecasting results, in order to select the most suited machine learning algorithms for the job.