target-leakage

⚠️ [已归档] 早期版本存在严重目标泄漏与真值定义错误，性能提升（偏差降低89.8%）为统计假象。请移步 dml-target-leakage-diagnosis 查看严谨诊断。

Human-in-the-loop Agentic ML audit system for tabular datasets — detects data risks, possible leakage, class imbalance, recommends metrics, benchmarks baseline models, tracks experiments, and generates grounded audit reports.

Reproducible NHANES codebase auditing whether CBC-based anemia classifiers retain hemoglobin-related target information after HGB removal through analyzer-equation reconstruction, controlled feature-ladder experiments, classifier comparisons, survey-aware inference, and out-of-cycle replication.

A controlled negative-result diagnostic study: how a single derived feature leaked the outcome Y(0) into DML covariates, appearing to cut ATE error by 89.8% while collapsing estimator variance 4x.

Customer churn analysis on 2,996 banking customers - and the target leakage that made the first result wrong

Which customers should a bank call? Campaign analysis in SQL, a leakage-aware response model, lift/gains analysis, and campaign ROI (Python, DuckDB)

Pre-training feature leakage auditor for tabular ML datasets. Checks column names, target correlation, categorical proxies, future timestamps, ID columns, and train/test distribution shift before any model is trained.

Privacy-conscious Polars and machine-learning methods for longitudinal membership analytics.

Evidence-based ML release-gate agent that catches cross-pipeline target and temporal leakage using DataHub column-level lineage