Evaluation, Leakage, and GDPR Boundaries

What This Is

This lesson teaches you how to distrust a flattering metric until the evaluation design has earned your trust.

Why This Matters in Production

Leaky evaluation produces false confidence, which is one of the fastest ways to launch a bad model with executive approval. Privacy mistakes add legal and reputational cost on top.

Mental Model

Evaluation is a claim about future usefulness. Leakage and privacy failures invalidate that claim by corrupting either the data boundary or the legal boundary.

Deep Dive

Leakage happens when information from outside the legitimate training context slips into features, preprocessing, or label construction. It often hides in time-aware data, aggregated statistics, or human-generated features. At the same time, privacy boundaries matter because model training can easily absorb identifiers that should have been removed, masked, or minimized. Maturity means treating metrics and privacy controls as one coherent quality system.

Worked Example

A support-ticket model uses a field that is only filled after escalation, but the target is escalation itself. Accuracy looks excellent. In reality, the model learned to detect a future artifact. That is leakage, not intelligence.

Common Failure Modes

Frequent failures include splitting after feature engineering, using test-set information in normalization, and assuming anonymization happened because someone said “the data is safe.”

official-doc

scikit-learn Metrics

Use official docs for metric definitions and tradeoffs.

law

GDPR Principles

Anchor privacy handling in a real legal principle.

official-doc

Data Leakage Guidance

Supplement the lesson with a practical framing of leakage risk.