Optimizing Service Cancellation Classifiers for Subscription-Based Online Video Platforms through Ensemble Learning

This research aims to optimize service cancellation classifiers for subscription-based online video platforms. It focuses on integrating self-reported user behavior data through ensemble learning techniques to mitigate the limitations of single-base models and enhance predictive robustness on complex and imbalanced datasets. The methodology encompasses data collection from 10,983 users, the use of the Synthetic Minority Over-sampling Technique (SMOTE) to address class imbalance, and a two-stage feature selection process using Boruta and Recursive Feature Elimination (RFE) to identify 10 critical predictors. The three primary determinants with the greatest influence on service cancellation are the number of subscribed channels (Subscribed), the volume of content consumed (ContentCount), and membership duration (Period). In terms of model performance, LightGBM emerged as the most effective single-base model both before and after hyperparameter tuning, yielding an accuracy of 0.8157. However, after ensemble learning frameworks were incorporated, the voting-based ensemble model increased overall accuracy to a maximum of 0.8204. Meanwhile, the mixed-based ensemble model, which uses logistic regression as the meta-model at the decision layer, demonstrated the most comprehensive ability to identify at-risk users, achieving the highest recall and Area Under the Receiver Operating Characteristic Curve (AUC-ROC) values of 0.8057 and 0.8833, respectively. Consequently, this research proposes a model-selection framework aligned with customer retention strategies: the mixed-based model is best suited to low-cost retention strategies that require comprehensive identification of at-risk users, whereas the voting-based model is appropriate for high-cost retention strategies that require high precision at the individual level, thereby enabling the cost-effective and sustainable management of marketing budgets and resources in the online media and entertainment platform industry.