A Hybrid Graph Neural Network and XGBoost Model for Telecom Customer Churn Prediction and Feature Attribution

Khan, Muhammad Osama Hassan (2025) A Hybrid Graph Neural Network and XGBoost Model for Telecom Customer Churn Prediction and Feature Attribution. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (3MB) | Preview
Preview	PDF (Configuration Manual) Download (809kB) | Preview

Abstract

This study investigates whether combining relational learning with high-performing tabular modelling improves telecom customer churn prediction while retaining practical explainability. Using a dataset of 7,043 subscribers with an imbalanced churn label, a customer graph was engineered from shared service-attribute relations and used to train a GraphSAGE encoder that generated 32-dimensional customer embeddings. Under controlled splits and matched metrics, tabular baselines (including standalone XGBoost) were compared with GraphSAGE-only learning and with two-stage hybrids that feed embeddings into downstream classifiers. Standalone XGBoost provided strong discrimination, while GraphSAGE alone achieved comparable ranking performance but displayed threshold sensitivity under class imbalance. The best results were obtained by the hybrid GraphSAGE+XGBoost pipeline, which substantially improved predictive performance (AUC-ROC ≈ 0.936, F1 ≈ 0.863) over the best tabular baseline. Explainability analysis linked churn risk to commercially meaningful drivers (tenure, contract, and charging variables) and incorporated graph-derived signals through embedding and structural features, allowing for both feature-level and relation-aware interpretation. Overall, the findings indicate that relational representations complement tabular predictors in churn modelling and that tree-boosted hybrids offer a practical route to deployable accuracy and interpretable retention insights. Results suggest prioritising validation-tuned thresholds and cost-sensitive targeting, while noting proxy edges limit causal claims and require privacy-aware governance procedures.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Simiscuka, Anderson UNSPECIFIED
Subjects:	H Social Sciences > HF Commerce > Marketing > Consumer Behaviour Q Science > Q Science (General) > Self-organizing systems. Conscious automata > Machine learning H Social Sciences > HD Industries. Land use. Labor > Specific Industries > Telecommunications Industry
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	08 Sep 2026 08:38
Last Modified:	08 Sep 2026 08:38
URI:	https://norma.ncirl.ie/id/eprint/9871

Actions (login required)

View Item