What question did this study set out to answer?

This study aims to address data leakage and enhance churn prediction accuracy through novel methodologies. It focuses on creating a reliable framework for telecom churn prediction.

April 28, 2026Open Access

A new perspective on data leakage prevention and adaptive attention in telecom churn prediction

Key Points

This study aims to address data leakage and enhance churn prediction accuracy through novel methodologies. It focuses on creating a reliable framework for telecom churn prediction.
Developed a leakage-free experimental protocol with Auto Balance for optimal sampling ratios.
Introduced LOP-Net architecture featuring Relationship-LSTM and LOP-Attention for capturing complex feature interactions.
Conducted experiments across multiple telecom datasets to validate the proposed methods.
Achieved accuracy of up to 96% and Brier scores near 3% using LOP-Net.
Statistical tests showed p-values below 5% when comparing LOP-Net with baseline models, indicating significant performance improvements.
Error-bar analysis confirmed the stability of performance across multiple random seeds.

Abstract

Customer churn prediction is a critical task in business analytics, as inaccurate forecasts can lead to unnecessary marketing expenditure and suboptimal retention strategies. Despite substantial progress in machine learning and deep learning, two fundamental challenges remain under-addressed: data leakage introduced during preprocessing and resampling, and the lack of architectures capable of modeling multi-level nonlinear feature interactions. Leakage frequently arises when scaling or oversampling is applied before data splitting, allowing information from validation or test partitions to influence model training. Meanwhile, existing architectures often depend on conventional module stacking without introducing genuinely new mechanisms for capturing hierarchical dependencies. This study makes two primary contributions. First, we establish a rigorous, leakage-free experimental protocol featuring Auto Balance, a standalone adaptive oversampling algorithm that automatically searches for optimal sampling ratios while strictly isolating validation and test sets. Auto Balance mitigates common leakage pathways in imbalanced classification and provides statistically reliable performance estimation. Second, we propose LOP-Net, a novel deep learning architecture incorporating two new modules—Relationship-LSTM, which captures ordered and unordered relational patterns among features, and LOP-Attention, which models multi-level, nonlinear interaction structures beyond standard attention mechanisms. Comprehensive experiments across multiple telecom datasets demonstrate that the combination of the leakage-free pipeline and LOP-Net consistently outperforms strong baselines, achieving accuracy of up to 96% and Brier scores near 3%. Performance robustness is further supported by error-bar analysis across multiple random seeds, demonstrating stable variance behavior, and by paired t-tests comparing LOP-Net with alternative models, where all p-values fall below 5%, confirming statistically significant performance improvements. An integrated interpretability dashboard additionally supports practical churn analysis and strategic decision-making.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper