Uplifting Model For Predicting Subscriber Churn Conversion Using Ensemble Learning: A Case Study Of Mobile Telecommunication Sector In Kenya

Abstract

Churn is the number one topic for Telco’s in Kenya and around the world. Customer churn in the telecommunication industry is still a big problem because emerging new technologies, lower costs, among other factors Churn brings with it many negative repercussions. While churn is a helpful key performance indicator for identifying areas of improvement whether in process or product, it can lead to financial disability eventually as customer acquisition cost are normally more astronomical than trying to please a disenchanted Subscriber. By analyzing churn drivers, we can safeguard the most import asset for a telecommunication company from churning. Predicting subscribers who are most likely to churn is fundamental for telecommunication companies. As a result, churn prediction is an important barometer for business success as well vastly studied and common activities that can be accomplished by machine learning applications for telecommunication industry. Telecommunication companies have since come to a realization that churn prediction only provide predictions but do not provide information for optimal decision making within a business setting. This is where uplift modelling has come to the fore. Uplift modeling is a branch of machine learning which aims at predicting the causal effect of an action such as a retention campaign or a marketing campaign on a given population by considering outcomes from the campaign treatment on that group, involving the sample populations that has been subjected to that campaign or treatment, and a control sample population. The model generated is then utilized to select the segment of population that the campaign would be profitable. This dissertation analyzes the use of ensemble methods in uplift modeling. The researcher will attempt to demonstrate higher performance compared to traditional classification and uplifting techniques. The researcher will attempt to show improved performance are a result of using ensemble classification techniques inculcating the differences in class probabilities in the treatment and control groups. The result being a Novel propensity outcome modification model. Safaricom plc was used as a case study to develop an uplifting for predicting subscriber churn conversion on various pre-existing subscriber segments. The objective of the study was to find the most profitable segment to target after using an ensemble classifier to predict probable churn customers on prepaid subscribers. Anonymized and pseudomized subscriber data was used for the study. The final results show the accuracy and precision of the ensemble predictive classifier and also the uplift scores for the various existing subscriber segments using the novel propensity outcome modification approach that identifies the probable segment to target with a retention campaign.