Answers are grounded in this paper and cite their sources.
We evaluate a hybrid framework for predicting next-month mobile number port-out using operational subscriber features, Word2Vec representations of service interactions, and Louvain community descriptors within XGBoost. A chronological protocol separates training (January–July 2025), probability calibration (August), threshold validation (September), and independent testing (October–November). The held-out cohort comprises 8,112,036 line-month records with 150,202 portability events (1.85%). Six models are compared under the same population and target. The full hybrid attains ROC-AUC 0.7266, PR-AUC 0.05215 and Brier score 0.017912, versus 0.7277, 0.05100 and 0.017922 for tabular XGBoost. XGBoost with Word2Vec leads PR-AUC (0.05298), whereas random forest leads ROC-AUC (0.7385). Class-weight analysis (1, 4 and 8) reveals a trade-off between recall and global discrimination. A frozen training-period graph contains 6704 small, highly fragmented communities, helping explain limited incremental graph value. We additionally report calibration, complete threshold-specific confusion matrices, provenance checks and a functional risk-analysis dashboard. The evidence supports representation-specific gains, not uniform superiority of combining all three methods.
No takes yet. Share an insight, caveat, or question.
Maguiña et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: