Course contentsShow
← Machine Learning and Deep Learning
Lesson 3208 of 3,53870. Interpretability: Feature-Based MethodsFree lesson

Desirable Properties of SHAP

Explore local accuracy, missingness, and consistency properties that make SHAP theoretically grounded.

Desirable Properties of SHAP

You'll learn why SHAP values are theoretically superior to other attribution methods through three key mathematical properties.

Why Properties Matter

Not all feature attribution methods are created equal. SHAP values satisfy three desirable properties that make them uniquely rigorous: local accuracy, missingness, and consistency. These aren't just nice-to-haves—they're mathematical guarantees that ensure your explanations behave intuitively and fairly.

The Three Pillars

Local Accuracy

The sum of SHAP values for all features plus a base value must exactly equal the model's prediction for that instance. Think of it like a budget: every contribution must account for the gap between the baseline prediction (what you'd predict without knowing anything) and the actual prediction. No unexplained residuals allowed.

Missingness

If a feature doesn't contribute to the model at all (like a constant column or an unused feature in a tree), its SHAP value must be exactly zero. This prevents spurious attributions—imagine crediting a disconnected wire for a lightbulb turning on. SHAP won't do that.

Consistency

If model A relies on a feature more than model B does (across all possible contexts), then that feature's SHAP value magnitude should be at least as large in model A. This ensures that when a feature becomes more important to the model's logic, its SHAP attribution reflects that increased importance.

Why This Matters

These properties distinguish SHAP from ad-hoc methods like simple gradients or coefficients. They guarantee your explanations decompose predictions faithfully, ignore irrelevant features, and respect changes in model behavior—making SHAP the gold standard for interpretability.

Key Takeaway: SHAP's local accuracy, missingness, and consistency properties provide mathematical guarantees that your feature attributions are faithful, meaningful, and theoretically sound.