Tag
提出在分布漂移下进行在线模型测试的方法学，通过区分人口变化与数据完整性问题两种漂移机制，在受控基线上复现漂移以估计模型当前性能，并评估修复数据完整性问题的潜在收益。
这篇论文对IBM Telco客户流失基准进行了可信度审计，揭示了报告高达99%准确率背后的四大问题：SMOTE前划分造成的数据泄漏、TotalCharges字段冗余性导致的SHAP解释失真、概率校准方法选择，以及F1最优阈值与成本最优阈值之间的巨大差距，并提出了一个可复现的报告检查清单。
This paper investigates how many days of early behavior are sufficient for subscription churn prediction, using the KKBox dataset. It shows that the optimal observation window depends on the experimental design, and that claims of window sufficiency should specify cohort construction, target definition, and feature families.
This paper proposes a rolling-window framework for customer churn prediction in non-contractual service environments, using 30-day behavioral windows to enable continuous risk assessment. Evaluated on real-world data, the feature-based model achieves 87.6% accuracy and 0.94 ROC-AUC, while the sequence-based model reaches 96.1% recall.
This paper evaluates traditional machine learning techniques (Random Forests, XGBoost, SVM) against a deep learning model (Unified Multi-Task Time Series Model) for customer churn prediction in retail, finding that conventional methods can outperform in predictive performance and efficiency.