Development of Churn Prediction Model using XGBoost - Telecommunication Industry in Sri Lanka

Maintaining a customer base at a feasible rate is considered important in most business organizations since customers are the precious asset of the business sector. It is a vital task to retain the customers at a steady level in any business enterprise for the overall stability of its business activities. Sufficient pieces of evidence in this connection have been gathered to prove that the telecommunication industry is the most affected field of business by the tendency of the customers to shift towards alternative service providers. Therefore, a distinctive effort has been made to design this specific forecasting method is carried out utilizing a combination of a properly approachable method aiming at clarifying the probability of the above-mentioned tendency of the clients seeking an alternative service provider in the industry. In this attempt, a data set that included 10, 000 postpaid consumer particulars including 20 attributes were taken for this research for a thorough analysis of this aggravating issue in the telecommunication industry. In the end, a satisfying outcome was witnessed and certain clarification was made out of the 10,000 subscribers 4888 showed positive attitudes and 5112 indicated negative to the churning behavior. Besides, this specific data set was subjected to complete verification in comparison to certain supervised machine learning algorithms such as Decision tree, Logistic Regression, Support Vector Machine (SVM), and Artificial Neural Networks (ANN). Along with this, ensemble techniques such as Random Forest, Extreme Gradient Boosting (XGBoost), and Adaptive Boosting (AdaBoost) also have been considered. Subsequently, an assurance was made that XGBoost possessed the ability to bring out the maximum and précised accuracy of 82.90%. Eventually, a hyperparameter tuning had been performed with XGBoost. As a result, an assurance was acquired that XGBoost showed an upsurge in the previously obtained accuracy from 82.90% to 83.13%.