auditing-train-test-split

Detect data leakage in your train-test split by auditing for identifier, target, and preprocessing leakage before trusting evaluation metrics. This process checks preprocessing order and group-aware partitioning to prevent inflated model performance.

Suspiciously high accuracy on time-series data often indicates temporal leakage during train-test splitting. Auditing temporal cutoffs and applying walk-forward or date-cutoff validation methods corrects this misleading evaluation.

The best way to split grouped or panel data without leakage is using group-aware partitioning. This approach ensures multiple rows belonging to the same person, customer, or session remain in the same split.

Validate train-test splits for imbalanced classification datasets by checking stratification logic. This ensures minority classes are proportionally represented across splits, preventing biased evaluation metrics.

Use stratified-group splitters instead of random splitting when your dataset contains both imbalanced classes and grouped entities. This method prevents leakage while maintaining class proportions across partitions.

Random splitting does not work for time-indexed datasets because it breaks temporal order and causes leakage. Auditing temporal cutoffs and applying time-series or walk-forward validation ensures realistic evaluation.