leakage-checklist
How to recognise data leakage: features unavailable at prediction time or proxies for the target.
- License 许可
- MIT
- 最近更新
- 2026-10-25
- Stars
- ★ 0
- Forks
- 0
01 前言 Front Matter
- name
- leakage-checklist
- metadata
- { "beyondml": { "kind": "knowledge" } }
- description
- How to recognise data leakage: features unavailable at prediction time or proxies for the target.
03 SKILL.md 正文
Leakage is a feature that would not exist when the prediction is made, or that is a proxy for the target.
- Recorded after the outcome: a churn model's Cancellation_Date, a fraud model's chargeback flag.
- Direct functions of the target: Tax_Amount for Price, total = sum of parts that include the target.
- Treatment implies diagnosis: a disease model's Prescription_A.
- IDs and timestamps that encode the split or the label order.
- Too good to be true: a single column that scores 0.98+ on its own almost always leaks.
Action: drop high-risk columns; for medium risk, keep them only with a written justification.
04 同仓库更多 Skills
tabular-demo
Demo step skill for tabular data.
leakage-demo
Demo knowledge skill about target leakage.
bad-name
Name does not match its folder.
train-test-split
How to split data so the score is honest.
tabular-features
Writes src/features.py with derived features (ratios, logs, interactions) and no target use. Usually right after tabular-eda. Check: run src/features.py, expects data/processed/features.csv.
tabular-export
Writes the shipped code (src/pipeline.py, src/train.py, src/predict.py) from the winning run and trains it. Usually right after evolve-search or tabular-ensemble. Check: run src/train.py, expects models/model.joblib, metric accuracy (classification) or r2 (regression) with a minimum.