| Item | Summary |
|---|---|
| Task | Clinical risk prediction / 临床风险预测 |
| Setting | WiDS Datathon 2026 competition workflow |
| Best public score | 0.97089 |
| Core methods | Gradient boosting ensembles, calibration, risk scaling |
| Project value | Competition ML, validation design, iterative experimentation |
This repository contains my competition workflow for WiDS Datathon 2026, focused on clinical risk prediction with gradient boosting ensembles, calibration, and leaderboard-oriented validation.
本仓库记录了我在 WiDS Datathon 2026 中的主要建模与实验流程，重点是临床预测任务中的集成学习、校准、特征工程和离线验证策略。
Goal / 目标
- Build a high-performing clinical prediction pipeline for the WiDS 2026 challenge.
- 构建一个高性能的临床预测系统，并尽可能缩小本地验证与 Kaggle leaderboard 之间的差距。
What makes this project interesting / 为什么这个项目值得看
- Not just a single model: this repo captures iterative experimentation across multiple ensemble strategies.
- 不只是一个模型，而是一整套从 baseline 到 blend / calibration / risk modeling 的实验演进。
- It reflects how I think about competition ML in practice: validation design, ablation, robustness, and leaderboard trade-offs.
- 它更像一个完整的竞赛研究工作区，而不是单次提交脚本。
- Public leaderboard score: 0.97089
- Primary modeling direction: gradient boosting ensembles + calibration
- Experiment themes: GBSA blend, subgroup odds scaling, rank calibration, piecewise gating, PLE-style exact blending
- Built and compared multiple post-ensemble calibration strategies instead of relying on a single leaderboard submission.
- 将集成之后的校准单独作为研究方向，而不是只做简单模型融合。
- Maintained a broad experiment log across multiple versions, including ablation, risk scaling, and blending studies.
- 保留了较完整的实验记录，能反映真实竞赛中的迭代过程。
src/ # Core modeling code
scripts/ # Experiment scripts and submission generation
experiments/ # Structured experiment subfolders
notebooks/ # Analysis notebooks
competition_analysis.md # Competition framing and observations
experiments.md # Experiment registry / notes
scripts/: experimental scripts for blending, calibration, and submission generationexperiments/: grouped experiment folders with results and notesnotebooks/: exploratory and validation notebooks
This project explores ensemble combinations built around strong gradient boosting baselines rather than treating a single model as final.
本项目核心不是“找一个最好模型”，而是围绕强基线构建更稳健的集成组合。
Several experiments focus on calibration-oriented techniques, including rank-based and quantile-based variants, to better align offline validation with leaderboard behavior.
多个实验围绕校准展开，目标是减少离线验证与 leaderboard 表现之间的偏移。
The project also explores subgroup-aware scaling and risk-sensitive adjustments, reflecting a more competition-realistic approach than pure average blending.
项目中还包含针对子群体与风险尺度的实验，这比单纯平均融合更贴近真实竞赛优化思路。
This repository reflects several practical lessons from competition work:
- Validation design matters as much as model architecture.
- 验证设计往往和模型本身一样重要。
- Small leaderboard gains often come from calibration and postprocessing, not just bigger models.
- 很多分数提升来自校准与后处理，而不只是更复杂的模型。
- A strong experiment log is essential for avoiding repeated mistakes.
- 清晰的实验记录能显著减少重复踩坑。
This repository is best viewed as a competition research workspace rather than a fully packaged benchmark repo.
这个仓库更适合被理解为一个竞赛研究工作区，而不是完全产品化、一步运行的标准 benchmark 仓库。
If I productionize it further, the next steps would be:
- add a cleaner entrypoint for training / inference
- normalize configs and paths
- add lightweight documentation for the main experiment families
- separate reusable utilities from one-off competition scripts
如果后续进一步整理，我会优先做：
- 统一训练 / 推理入口
- 整理配置与路径
- 为主要实验方向补充更清晰的文档
- 把可复用模块和一次性竞赛脚本彻底拆开
This project is one of the best examples of how I work in machine learning competitions:
- iterative modeling
- systematic ablation
- leaderboard-aware validation
- practical engineering under uncertainty
这个仓库最能体现我在 ML 竞赛里的工作方式：
- 持续迭代
- 系统化实验
- 面向 leaderboard 的验证设计
- 在不确定条件下推进工程实现