Telecom Churn Prediction: An approach Towards Big Data

• Autor(a) principal: Coelho, António Fonseca
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10362/160464
• topic: Churn; Classification; Modeling; Machine learning; Telecom; Domínio/Área Científica::Engenharia e Tecnologia::Engenharia Eletrotécnica, Eletrónica e Informática

Churn prediction is a crucial subject in telecom companies. Acquiring a new customer is more expensive than retaining a customer. Identifying such customers requires the address of multiple challenges. The first is caused by telecom datasets. These tend to be high-dimensional and at the same time very sparse, bringing multicollinearity and overfitting issues. Another challenge concerns variable data types. There are static variables and dynamic variables over time. The nature of the use case creates another adversity. The goal is to predict who is leaving the service, but, in any successful company, there are much more clients staying than leaving. This creates what an unbalanced dataset, where the binary target variable has an unbalanced distribution between its’ classes. In this work, a pipeline is proposed targeting the telecom industry. This pipeline aims to address the churn problem, i.e., to identify the clients that have a high propensity to leave the service. The pipeline is designed to deal with the multiple challenges identified and to be adaptable to other telecom datasets. This pipeline is composed of multiple steps, the first step was to restructure data, this was done by realigning all clients by its last month, in an active state, stored in the system. The multiple observations per client were compressed into one using statistics like median and standard deviation, after that feature selection method was applied but multiple options were considered and evaluated at the end of this document. Models were then used to predict variable. These models were adapted to handle unbalance challenge. This work demonstrated the ability to achieve reasonable results using a restructuring proccess and compressing statistics. This work also demonstrated the ability to achieve reasonable good results using a feature selection algorithm.