Machine Learning-Based Customer Churn Prediction in Subscription Publishing utilizing CRISP-DM methodology: An Automated Pipeline for Multi-Publisher Environments

In the subscription-based publishing industry, customer churn poses a significant challenge to business sustainability, as acquiring new customers is substantially more costly than retaining existing ones. This study examines the development of an automated churn prediction pipeline for S.P. AbonneeService, a B2B subscription service provider managing over 200 titles and 350,000 end-consumers across multiple publishing categories. The research implements a comprehensive machine learning framework utilizing the CRISP-DM methodology, evaluating six algorithms (Naive Bayes, Logistic Regression, Random Forest, XGBoost, LightGBM, SVM) across three resampling techniques and three temporal validation strategies using five years of historical subscription data from three distinct publishing companies. The automated preprocessing pipeline addresses heterogeneous data structures, seasonal variance, and class imbalance through systematic feature engineering, temporal validation, and synthetic minority oversampling. Experimental results demonstrate that LightGBM with SMOTE resampling achieves superior performance across all evaluated contexts, with AUC-PR values exceeding 0.95 and precision rates above 0.95 for top-performing configurations. The study establishes that automated churn prediction systems can deliver exceptional predictive performance while maintaining interpretability essential for actionable retention strategies, enabling subscription publishing companies to implement advanced predictive capabilities that directly support customer retention.