GRAPH NEURAL NETWORKS FOR CHURN CONTAGION PREDICTION IN TELECOMMUNICATIONS: A RELATIONAL LEARNING FRAMEWORK FOR CUSTOMER NETWORK ANALYSIS

Abstract

Customer churn prediction in telecommunications has traditionally treated subscribers as independent entities, ignoring the relational structures—family plans, corporate accounts, and social referral networks—that govern collective departure behaviour. This article proposes a Graph Neural Network (GNN)–based framework that models churn as a network contagion phenomenon, leveraging GraphSAGE convolutional layers to propagate behavioural signals across customer relationship graphs. Using a dataset of 65,000 subscriber records enriched with inferred relational edges (shared billing, call-pattern reciprocity, and temporal co-activity), the proposed model captures churn cascades that conventional classifiers miss. Experimental results demonstrate that the GNN framework improves recall of high-value churners by 18.3 percentage points over a hybrid CNN-LSTM baseline, while reducing false-negative revenue leakage by an estimated ₦2.1 million per quarter. Explainability is preserved through attention-weighted edge analysis, revealing that churn contagion propagates most aggressively through corporate account hierarchies and family-plan primary holders. The findings establish relational modelling as a critical frontier in telecom retention analytics and provide a deployable architecture for proactive, network-aware intervention strategies.