Telecom churn – reproducibility package (code, calibrated scores, models, figures)

This deposit contains code and artifacts for a fully reproducible churn-prediction workflow on IBM’s Telco Customer Churn dataset. The pipeline emphasizes class-imbalance aware model selection (PR-AUC with class weights), probability calibration (isotonic; reported via Brier score and reliability curves), and transparent explainability (global SHAP + local LIME). Calibrated probabilities are mapped to operational risk tiers (Low <0.35, Medium 0.35–0.59, High ≥0.60) to support budgeted retention actions. Contents: (i) training/evaluation scripts, (ii) a calibrated model artifact and evaluation figures (PR/ROC, gains, calibration), (iii) LIME figures and selection logs under a deterministic Plan-A protocol (Top-risk / Median-of-tier / Near-threshold per tier), (iv) audit reports and a Zenodo-ready README. Random seed is fixed to 42 for repeatability. Data availability: we do not redistribute the raw IBM/Kaggle CSV. Users should download WA_Fn-UseC_-Telco-Customer-Churn.csv from the original source and place it in the project root as documented in the README.本存档包包含用于在IBM电信客户流失（Telco Customer Churn）数据集上搭建全流程可复现客户流失预测工作流的代码与模型制品。该工作流重点关注考虑类别不平衡的模型选择（结合类别权重的精确率-召回率曲线下面积（PR-AUC）指标）、概率校准（采用保序回归方法，通过布里尔分数（Brier score）与可靠性曲线进行性能评估）以及可解释性分析（全局SHapley Additive exPlanations（SHAP）与局部Local Interpretable Model-agnostic Explanations（LIME）方法）。校准后的预测概率将被映射至运营风险层级（低风险<0.35、中风险0.35–0.59、高风险≥0.60），以支撑预算规划内的客户留存举措。 存档包内容包括：(i) 训练与评估脚本；(ii) 校准后的模型制品与评估图表（包含PR/ROC曲线、增益曲线与校准结果可视化图）；(iii) 遵循确定性Plan-A协议（按各风险层级分别选取高风险样本、层级中位数风险样本与临界值附近样本）的LIME可视化图表与模型选择日志；(iv) 审计报告与可提交至Zenodo平台的项目自述文件。为确保实验可复现，随机种子固定为42。 数据可用性说明：本存档不附带IBM/Kaggle平台的原始CSV数据集。使用者需从原始来源下载WA_Fn-UseC_-Telco-Customer-Churn.csv，并按照自述文件中的指引将其放置于项目根目录下。