Unveiling the Power of Social Influence: A Machine Learning Framework for Churn Prediction With Network Analysis

Abstract

Customer churn is a significant concern for firms due to the high cost of acquiring new customers. The expenditure related to developing new consumers surpasses that of customer retention. Customer churn prediction models were given to analyze the impact of this problem on organizations' revenues. These models primarily utilize machine learning algorithms to predict outcomes using data from demographic factors and customer service information components. This study investigates the impact of social relationships on customer churn probability and evaluates the performance of machine learning methods after introducing a new concept called the conformity factor. To improve the performance of standard machine learning models, we performed feature engineering by leveraging phone call network data and developing influence and conformity metrics. These metrics capture the social connections of individuals within the network. We employed various machine learning classification approaches and evaluated their performance using standard measures like AUC, accuracy, precision, F1-score, MCC, Cohen's kappa, and Brier score. The experiments demonstrated that incorporating these social network variables, particularly the proposed influence and conformity indices, significantly enhanced the performance of all churn prediction models developed in this study. Among the tested approaches, the gradient boosting model achieved the highest level of performance.