Scholarlink

<p>Customer churn prediction is critical for subscription-based businesses seeking to implement effective retention strategies. However, real-world churn datasets are often highly imbalanced and contain large volumes of unlabeled data, necessitating algorithms that can handle both semi-supervised and class-imbalanced learning. Safe semi-supervised learning algorithms develop semi-supervised models without performance degradation. This work introduces CISL, a class-imbalanced safe semi-supervised learning approach designed to learn robustly from such challenging data. To our knowledge, this is the first application of safe semi-supervised learning to churn prediction and among the first to address class imbalance within this paradigm. CISL employs a three-phase pseudo-labeling strategy with integrated safety mechanisms: (1) SSL-guided pseudo-label assignment, (2) confidence-based class-balanced batch selection, and (3) performance-driven safety verification. Experiments on four real-world churn datasets across six labeling ratios show that CISL consistently outperforms competitive methods, achieving average improvement percentages of 3.6%, 2.99%, and 2.65% in Geometric Mean, F1-measure, and Matthews Correlation Coefficient, respectively. These improvements are statistically validated using Friedman ranking and Wilcoxon post hoc tests, and CISL demonstrates superior safety adherence compared to baseline methods.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Class imbalanced safe semi-supervised learning for Customer Churn Prediction (CISL)

  • 期刊论文

  • 发表日期：

    2026年6月19日

查看全文
  • Salma Abdelmonem,
  • Dina Elreedy,
  • Samir I. Shaheen

摘要

Customer churn prediction is critical for subscription-based businesses seeking to implement effective retention strategies. However, real-world churn datasets are often highly imbalanced and contain large volumes of unlabeled data, necessitating algorithms that can handle both semi-supervised and class-imbalanced learning. Safe semi-supervised learning algorithms develop semi-supervised models without performance degradation. This work introduces CISL, a class-imbalanced safe semi-supervised learning approach designed to learn robustly from such challenging data. To our knowledge, this is the first application of safe semi-supervised learning to churn prediction and among the first to address class imbalance within this paradigm. CISL employs a three-phase pseudo-labeling strategy with integrated safety mechanisms: (1) SSL-guided pseudo-label assignment, (2) confidence-based class-balanced batch selection, and (3) performance-driven safety verification. Experiments on four real-world churn datasets across six labeling ratios show that CISL consistently outperforms competitive methods, achieving average improvement percentages of 3.6%, 2.99%, and 2.65% in Geometric Mean, F1-measure, and Matthews Correlation Coefficient, respectively. These improvements are statistically validated using Friedman ranking and Wilcoxon post hoc tests, and CISL demonstrates superior safety adherence compared to baseline methods.