ML//Evaluation//data leakage

When information the model would not have at prediction time sneaks into training, inflating scores that then collapse in the real world.

When information the model would not have at prediction time sneaks into training, inflating scores that then collapse in the real world.

Does not need to look like cheating: a clean random split can leak when near-duplicate rows (same session, adjacent time windows) land on both sides.

The reason to split by group (session, patient, user), not by row. See cross-validation.

A post-hoc feature not available at prediction time is a classic leak.