Significance-Based Feature Extraction for Customer Churn Prediction Data in the Telecom Sector

Abstract

The telecom industry is saturated with many service providers competing for highly rational customers. The current big data and highly technological era calls for real-time churn analysis and decision making which has also been highlighted in previous studies. However, telecom data is highly dimensional in nature thus when this is coupled with this big data era increases the computational and processing costs. Therefore, this complexity and dimensionality of telecom data coupled with the current need for near or real-time churn analysis demands feature selection-based models that efficiently consider the most relevant variables in explaining customer churn behaviors. This study proposes a feature extraction-based churn prediction model that concentrates on the most relevant features with significant discriminatory power for churn. The data has been reduced on the basis of missing values and irrelevant variables. Irrelevant variables were first identified by use of Random Forest and Logistic Regression models. The findings of the study provide churn analysts with insights about the prediction errors to consider and minimize in their future churn analyses. It also contributes to reducing computational costs incurred by churn analysts working with big data in their churn prediction and analysis.