Data leakage is the unintended transfer of information from a test or deployment context into a training context, causing a model to learn signals that are unavailable at prediction time. It occurs when preprocessing, feature selection, or any data transformation accesses information from outside the training fold — for example, scaling features using the full-dataset mean, or selecting features ranked by correlation with the target computed over all samples. Its persistence mechanism is cultural and procedural: it is catalogued as a failure mode in ML literature, taught as a cautionary principle in coursework, and mitigated through protocol design such as cross-validation pipelines and strict data-folding discipline. [formal: fuga datorum | substrate: mind | horizon: a life | explicit: yes | epoch: 0.02]
Accepted ontology entry
data leakage
Data leakage is the unintended transfer of information from a test or deployment context into a training context, causing a model to learn signals that are unavailable at prediction time. It occurs when preprocessing, feature selection, or…
Definition
Why it is in scope
A human-made concept in machine learning methodology describing the unintended information flow from training data into model evaluation, violating the boundary between training and testing phases. It persists through shared terminology in ML literature and practice, as a failure mode that practitioners must actively prevent.
Names and aliases
- data leakageen · CANONICAL
Relations from this entry
- cmsernc8k05tg3vv3qo17g356INSTANCE_OF →
Data leakage IS a specific kind of leakage (information escaping unintended channels). In ML, it specifically describes information flow from training to evaluation data — a subtype of the broader leakage construct.
- cmsegyrt305g13vv3pa46s19tDEPENDS_ON →
Data leakage as a concept only operates within a train-test paradigm: it describes information from training data improperly reaching the evaluation pipeline. Remove the train-test split and the concept of data leakage stops operating entirely — it is defined by the relationship between training and evaluation data. Per Law 8, this is present-tense dependency, not historical association.
Relations to this entry
No accepted relations in this direction.
Record identity
- Created
- Aug 4, 2026, 3:46 PM UTC
- Content hash
- 62a50e36ab8d3b18f8af9ef9fba4a9cf23f09dcebab53c90e8cf4febdce34f91