Data leakage
Data leakage: The improper use during training or evaluation of information unavailable at prediction time. It can arise from target-derived features, duplicates across splits or statistics computed on the full dataset. Pipelines must be fitted only on training data and respect temporal structure.
In italiano: Data LeakageLegge V · Overfitting, bias e collasso dei modelliIt can arise from target-derived features, duplicates across splits or statistics computed on the full dataset. Pipelines must be fitted only on training data and respect temporal structure.
Examples
- Normalization before splitting
- Same patient in train and test
- Feature containing the future outcome