A Survey of Evolution in Predictive Models and Impacting Factors in Customer Churn

Abstract

The information-based prediction models using machine learning techniques have gained massive popularity during the last few decades. Such models have been applied in a number of domains such as medical diagnosis, crime prediction, movies rating, etc. Similar is the trend in telecom industry where prediction models have been applied to predict the dissatisfied customers who are likely to change the service provider. Due to immense financial cost of customer churn in telecom, the companies from all over the world have analyzed various factors (such as call cost, call quality, customer service response time, etc.) using several learners such as decision trees, support vector machines, neural networks, probabilistic models such as Bayes, etc. This paper presents a detailed survey of models from 2000 to 2015 describing the datasets used in churn prediction, impacting features in those datasets and classifiers that are used to implement prediction model. A total of 48 studies related to churn prediction in telecom industry are discussed using 23 datasets (3 public and 20 private). Our survey aims to highlight the evolution of techniques from simple features/learners to more complex learners and feature engineering or sampling techniques. We also give an overview of the current challenges in churn prediction and suggest solutions to resolve them. This paper will allow researchers such as data analysts in general and telecom operators in particular to choose best suited techniques and features to prepare their churn prediction models.