Scholarlink

<p>We present a profit-weighted ensemble learning framework for customer churn prediction that integrates customer economic value throughout the modeling pipeline. Our approach incorporates value considerations at three levels: base model training through profit-weighted sampling, ensemble construction via value-aware meta-learning, and decision threshold optimization. We introduce value-weighted AUC and top decile lift as novel evaluation metrics that prioritize high-value customer classification accuracy. Through comprehensive evaluation across banking, e-commerce, and streaming domains using nested cross-validation, we compare our method against established profit-oriented approaches. Results reveal a fundamental trade-off between discrimination capability and profit optimization: while our ensemble achieves competitive value-weighted AUC performance, segmentation-based methods consistently outperform continuous value-weighting approaches in economic outcomes across all domains. Cross-domain analysis demonstrates that customer heterogeneity represents the primary driver of profit optimization in churn prediction, with explicit segmentation strategies proving more effective than algorithmic sophistication alone. Our methodological contributions provide novel business-aligned metrics and evaluation tools for future research in profit-oriented machine learning applications.</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

Profit-Weighted Machine Learning: A Cross-Domain Analysis of Ensemble Approaches to Churn Prediction

  • 期刊论文

  • 发表日期：

    2026年3月20日

查看全文
  • Abdelmounaim Kerkri

摘要

We present a profit-weighted ensemble learning framework for customer churn prediction that integrates customer economic value throughout the modeling pipeline. Our approach incorporates value considerations at three levels: base model training through profit-weighted sampling, ensemble construction via value-aware meta-learning, and decision threshold optimization. We introduce value-weighted AUC and top decile lift as novel evaluation metrics that prioritize high-value customer classification accuracy. Through comprehensive evaluation across banking, e-commerce, and streaming domains using nested cross-validation, we compare our method against established profit-oriented approaches. Results reveal a fundamental trade-off between discrimination capability and profit optimization: while our ensemble achieves competitive value-weighted AUC performance, segmentation-based methods consistently outperform continuous value-weighting approaches in economic outcomes across all domains. Cross-domain analysis demonstrates that customer heterogeneity represents the primary driver of profit optimization in churn prediction, with explicit segmentation strategies proving more effective than algorithmic sophistication alone. Our methodological contributions provide novel business-aligned metrics and evaluation tools for future research in profit-oriented machine learning applications.