Predicting and Reducing Wireless Customer Churn Using the AI2-Sriya Expert Index (SXI)

Abstract

The telecommunication sector has experienced swift technological development, with wireless services being the forces of change. One of the key issues in this sector is customer churn that affects profitability and market competitiveness. As customer expectations increase and competition becomes fiercer, telecom providers should be concerned with retaining current customers through better services. This study examines churn prediction in the wireless telecommunication sector via machine learning (ML) methods, focusing on the performance of the AI2-Sriya Expert Index (SXI) model. AI2(AI Square), launched by Sriya.AI, improves standard AI with a second layer of AI to facilitate adaptive learning. Fundamentally, SXI is a super feature that aggregates outputs from 5–10 ML models into one dynamic score that accurately determines the probability of customer churn. This method condenses complex, multi-dimensional data into two-dimensional data for enhanced decision-making and predictive analysis. In the comparison, conventional models such as XGBoost had 67% accuracy and 0.68 AUC, other research employing LightGBM and CatBoost had AUC values around 0.9, and internal benchmarking showed TabNet and FT-Transformer achieving AUCs of 0.65 and 0.66, respectively. CatBoost with hyperparameter optimization obtained an F1-score of 93% and AUC of 0.91. Nevertheless, the SXI model outperformed all of them with 98% accuracy and a 0.98 AUC value. Aside from predictive accuracy, SXI provides actionable business insights through iterative optimization with deep learning. This study shows that SXI has the capability to decrease customer churn from 49.5% to 39.65% (first 20% improvement). Through optimization, churn can be further reduced to 24.78% (50% reduction) and then eventually to 6.44% (87% reduction), providing long-term strategic advantages. These findings validate that the SXI framework offers a robust, precise, and interpretable solution for customer churn in telecom that is substantially better than that of conventional ML models.