Label Leakage
The inadvertent inclusion of output information in training data labels, which can inflate performance metrics and conceal true model generalization issues.
Occurs when features inadvertently encode the target label (e.g., timestamp or ID that correlates perfectly), causing the model to “cheat” rather than learn real patterns. Label leakage leads to overly optimistic validation scores and subsequent production failure. Governance demands rigorous feature-label correlation analysis, hold-out test sets from different timeframes, and pipeline checks to prevent leakage in feature engineering.
A churn-prediction dataset includes a “cancellation_reason” column labeled only after the customer left, perfectly predicting churn. After discovering this leakage, the team removes the column, retrains with only pre-cancellation features, and validates performance on an entirely new cohort - revealing the true predictive power.
We help you find answers
What problem does Enzai solve?
Enzai provides enterprise-grade infrastructure to manage AI risk and compliance. It creates a centralized system of record where AI systems, models, datasets, and governance decisions are documented, assessed, and auditable.
Who is Enzai built for?
How is Enzai different from other governance tools?
Can we start if we have no existing AI governance process?
Does AI governance slow down innovation?
How does Enzai stay aligned with evolving AI regulations?
Research, insights, and updates
Empower your organization to adopt, govern, and monitor AI with enterprise-grade confidence. Built for regulated organizations operating at scale.