alohe15/mle-portfolio

# MLE 作品集 用于 IEEE-CIS 欺诈检测的端到端机器学习项目：包含数据准备、特征工程、模型训练、评估和提供服务的 API。 ## 项目结构 ``` mle-portfolio/ ├── configs/ # Hyperparameters, feature lists, paths (edit settings without code changes) ├── data/ │ ├── raw/ # Competition CSVs and archives (gitignored) │ └── processed/ # Merged and featured parquet files (gitignored) ├── docs/ # Session logs and write-ups ├── evals/ # Cross-version comparisons and evaluation artifacts ├── models/ # Trained model files and per-run metrics (gitignored) ├── notebooks/ # Exploratory analysis scripts ├── scripts/ # Pipeline: merge → features → train ├── services/ # FastAPI inference service ├── tests/ # API integration and latency checks └── requirements.txt # Pinned dependencies ``` ## 环境设置 ``` python -m venv .venv source .venv/bin/activate pip install -r requirements.txt ``` 将原始的 IEEE-CIS 数据放在 `data/raw/` 目录下（参见 Kaggle 竞赛文件）。 ## Pipeline 在仓库根目录下运行： ``` python scripts/merge_train_data.py python scripts/feature_engineering.py python scripts/train_baseline_lightgbm.py python scripts/train_engineered_lightgbm.py python evals/compare_models.py ``` ## 提供预测服务 ``` uvicorn services.api.app:app --reload --app-dir . python tests/test_api_endpoint.py ``` ## 配置 训练超参数、特征工程设置和文件路径位于 `configs/` 目录下： - `configs/training.json` — 划分比例、决策阈值、LightGBM 参数 - `configs/feature_engineering.json` — 空值阈值、频率编码列 - `configs/paths.json` — 数据和模型产物位置

标签：Apex, AV绕过, FastAPI, LightGBM, 反欺诈检测, 数据工程, 机器学习, 模型部署, 逆向工具