Telecom Customer Churn Identification Using AdaBoost Algorithm in Subscription-Based Service Models

Customer churn—the voluntary discontinuation of a service subscription—represents the single largest controllable revenue loss category for subscription-based telecom operators, with mean annual churn rates of 15-25% across mobile, broadband, and converged service providers globally in 2024. The asymmetry between subscriber acquisition cost (mean USD 284 per retained customer versus USD 1,240 per newly acquired equivalent) creates a strong economic case for predictive churn identification systems capable of triggering personalised retention interventions before a churn decision is made. This paper evaluates the AdaBoost (Adaptive Boosting) ensemble learning algorithm for telecom customer churn prediction, benchmarked against four alternative models on a proprietary dataset of 148,400 subscriber records from three telecommunications operators across two countries, covering 18 months of behavioural data (January 2023-June 2024). Fortytwo engineered features spanning six behavioural and transactional categories are constructed from raw billing, network quality, usage, and customer service interaction records. AdaBoost (250 weak learners; base estimator: decision stump, max depth 1; learning rate: 0.8; SAMME.R algorithm) achieves$\mathbf{F} \mathbf{1}=\mathbf{0. 8 4 7}$, precision$=\mathbf{0. 8 3 1}$, recall$=\mathbf{0. 8 6 4}$, and AUC-$\mathbf{R O C}=\mathbf{0. 9 1 2}$on the held-out test set (20% stratified split,$\mathbf{n}=\mathbf{2 9, 6 8 0}$), outperforming Logistic Regression (F1: 0.724), Decision Tree (F1: 0.748), Random Forest (F1: 0.812), and XGBoost (F1: 0.838). SHAP feature attribution identifies contract tenure (24.2% of model output variance), monthly charge-to-plan-value ratio (18.6%), customer support call frequency (16.4%), network quality score (14.8%), and days since last plan upgrade (11.2%) as the five dominant churn predictors, collectively contributing 85.2% of total SHAP variance. Deployed as a monthly scoring pipeline across the three operator subscriber bases, the AdaBoost churn model supports a targeted retention intervention programme that reduces the observed churn rate from 18.4% to 11.2% over the 18-month evaluation window—a 39.1% relative reduction—with an estimated aggregate annual revenue preservation of USD 4.8 million and a customer lifetime value improvement of 24.6%.

Hali tarjima qilinmagan