Atharva-khetale/churn-mlops

# 🔁 流失预测 MLOps 流水线 [](https://github.com/YOUR_USERNAME/churn-mlops/actions) [](https://codecov.io/gh/YOUR_USERNAME/churn-mlops) [](https://python.org) [](https://mlflow.org) [](https://fastapi.tiangolo.com) [](https://docker.com) [](https://dvc.org) [](https://YOUR_API_URL.onrender.com/docs) ## 🏗️ 架构 ``` Git Push │ ▼ GitHub Actions CI/CD ├── ruff + black (code quality) ├── pytest (45 tests, 82% coverage) ├── python pipeline.py (train all models) └── Docker build + smoke test │ ▼ StreamLit Deploy │ ▼ ┌─────────────────────────────────────────────┐ │ MLOps Pipeline │ │ │ │ CSV ──► Preprocessing ──► Training │ │ ├── XGBoost │ │ ├── RF │ │ └── GBM │ │ │ │ │ MLflow Track │ │ + Registry │ │ │ │ │ best_model.pkl │ │ │ │ │ FastAPI Serve │ │ /predict /metrics │ │ │ │ │ Drift Monitor (PSI) │ │ Weekly Retrain Cron │ └─────────────────────────────────────────────┘ ``` ## 🛠️ 技术栈 | 层级 | 技术 | |-------|-----------| | 语言 | Python 3.11 | | ML 模型 | XGBoost, Random Forest, Gradient Boosting | | 实验跟踪 | MLflow (参数, 指标, artifacts, model registry) | | API 服务 | FastAPI + Uvicorn | | 容器化 | Docker (多阶段构建) | | CI/CD | GitHub Actions | | 代码质量 | ruff + black | | 测试 | pytest + pytest-cov (82% 覆盖率) | | 数据版本控制 | DVC | | Drift 检测 | PSI + Evidently AI | | 部署 | Render / Railway / Fly.io | | 数据集 | IBM Telco Customer Churn (7,043 行, 20 个特征) | ## 📁 项目结构 ``` churn-mlops/ ├── src/ │ ├── data_ingestion.py # Download IBM dataset, schema validation │ ├── preprocessing.py # LabelEncoder, StandardScaler, train/test split │ └── train.py # MLflow multi-model training + business metrics │ ├── api/ │ └── main.py # FastAPI: /predict /batch /metrics /health /logs │ ├── monitoring/ │ └── drift_monitor.py # PSI + Evidently + prediction log analysis │ ├── tests/ │ ├── conftest.py # Shared fixtures │ └── test_pipeline.py # 45 tests: data, preprocessing, model, API, drift │ ├── .github/workflows/ │ ├── mlops_ci.yml # CI/CD pipeline │ └── retrain.yml # Weekly automated retraining │ ├── pipeline.py # End-to-end orchestrator ├── Dockerfile # Multi-stage production container ├── render.yaml # Render deployment config ├── dvc.yaml # DVC pipeline stages ├── pyproject.toml # Tool configuration ├── HUMAN_CHANGES_AND_SETUP.md # Full setup guide ← read this first └── requirements.txt ``` ## 🚀 快速开始 ``` git clone https://github.com/YOUR_USERNAME/churn-mlops.git cd churn-mlops pip install -r requirements.txt # 完整 pipeline (ingest → preprocess → train → monitor) python pipeline.py # MLflow UI mlflow ui --port 5000 # 启动 API uvicorn api.main:app --reload --port 8000 # 测试 + coverage pytest tests/ -v --cov=src --cov=api --cov=monitoring # 简单方式 (将运行 frontend+Backend) python streamlit_app.py ``` ## 📊 MLflow 实验 每次 3 个模型运行都会跟踪： **ML 指标** - `accuracy`, `f1_score`, `roc_auc`, `precision`, `recall` **业务指标** ← _这是它的突出之处_ - `business/revenue_saved_usd` — 捕获的预测流失客户 × 留存 ROI - `business/wasted_spend_usd` — 假阳性 × 留存成本 - `business/net_impact_usd` — 挽回的收入减去浪费的支出 - `business/missed_revenue_usd` — 假阴性 × 生命周期价值 **Artifacts** - 混淆矩阵、ROC 曲线、特征重要性图表 ## 🌐 API Endpoints | 方法 | Endpoint | 描述 | |--------|----------|-------------| | `GET` | `/` | 欢迎 + 文档链接 | | `GET` | `/health` | 模型存活检查 | | `POST` | `/predict` | 单个客户预测 | | `POST` | `/predict/batch` | 批量预测（最多 500 个） | | `GET` | `/metrics` | 聚合预测统计信息 | | `GET` | `/logs?last=50` | 最近的预测日志 | | `POST` | `/reload` | 热重载模型 artifacts | ### 请求示例 ``` curl -X POST https://YOUR_API_URL.onrender.com/predict \ -H "Content-Type: application/json" \ -d '{ "gender": "Female", "SeniorCitizen": 0, "Partner": "Yes", "Dependents": "No", "tenure": 6, "PhoneService": "Yes", "MultipleLines": "No", "InternetService": "Fiber optic", "OnlineSecurity": "No", "OnlineBackup": "No", "DeviceProtection": "No", "TechSupport": "No", "StreamingTV": "No", "StreamingMovies": "No", "Contract": "Month-to-month", "PaperlessBilling": "Yes", "PaymentMethod": "Electronic check", "MonthlyCharges": 70.70, "TotalCharges": 151.65 }' ``` ### 响应示例 ``` { "churn_probability": 0.8234, "churn_prediction": true, "risk_level": "HIGH", "expected_revenue_loss": 645.42, "timestamp": "2024-06-01T10:30:00Z", "model_version": "2.0.0", "latency_ms": 4.2 } ``` ## 🧪 测试 ``` pytest tests/ -v 45 passed in 26s Coverage: src/train.py 91% api/main.py 91% src/preprocessing.py 80% monitoring/... 64% TOTAL 82% ``` 测试涵盖： - 数据接入 + schema 验证 - 预处理（清洗、编码、缩放、拆分） - 模型训练（概率、AUC、业务指标） - 图表生成（混淆矩阵、ROC、特征重要性） - 完整的 MLflow `train_and_track` 集成 - Drift 监控（PSI、检测阈值） - 所有 10 个 FastAPI endpoints，包括 503 错误处理 ## 🔄 CI/CD 流程 ``` git push main │ ├── ruff check (linting) ├── black --check (formatting) │ ├── pytest (45 tests, 82% coverage) │ ├── python pipeline.py (train + MLflow) │ ├── docker build + smoke test │ ├── GET /health → 200 ✅ │ └── POST /predict → probability ✅ │ └── Render auto-deploy → live URL ``` **每周（每周日 UTC 02:00）：** ``` GitHub Actions cron └── python pipeline.py → retrain on latest data ``` ## 📈 模型性能 | 模型 | ROC-AUC | F1 | Precision | Recall | |-------|---------|-----|-----------|--------| | XGBoost | ~0.84 | ~0.62 | ~0.68 | ~0.57 | | Random Forest | ~0.82 | ~0.60 | ~0.65 | ~0.55 | | Gradient Boosting | ~0.83 | ~0.61 | ~0.66 | ~0.56 | _由于采样中的随机性，每次运行的指标会略有不同._ ## 🗃️ 数据集 **IBM Telco Customer Churn** (开源) - 记录：7,043 名客户 - 特征：20 个（人口统计、服务、账单） - 目标：Churn（是/否）— 26.5% 流失率 - 来源：[IBM GitHub](https://github.com/IBM/telco-customer-churn-on-icp4d) ## 📖 设置指南 请参阅 **[HUMAN_CHANGES_AND_SETUP.md](HUMAN_CHANGES_AND_SETUP.md)** 了解： - 完整的本地设置说明 - 如何进行公开部署 - GitHub secrets 配置 - DVC 数据版本控制设置 - 可以在简历上写些什么 ## 👤 作者 作品集 MLOps 项目 — 展示生产级 ML 工程技能。 ⭐ 如果这个仓库对您有帮助，请点个 Star！

标签：Apex, AV绕过, DVC, FastAPI, Kubernetes, MLOps, 安全规则引擎, 开源框架, 持续集成, 机器学习, 模型部署, 请求拦截, 逆向工具