Phase 3: Explainability & Transparency · 50 min · Python · SHAP · scikit-learn

SHAP — From Global to Local Explanations

Every prediction has a story. SHAP tells it with math.

What you will learn

Explain Shapley values and their origin in cooperative game theory
Use TreeExplainer for tree-based models and KernelExplainer for model-agnostic explanations
Generate SHAP global summary plots and local waterfall/force plots
Interpret SHAP values for technical and non-technical stakeholders

Objective

Test whether an explanation is faithful, understandable, decision-relevant, and appropriate for its actual audience.

The Problem

You have a trained model. It's accurate. It's also a black box. You need to answer two questions:

Global: Which features matter most across all predictions?
Local: Why was this specific person denied a loan?

SHAP (SHapley Additive exPlanations) answers both. It's based on Shapley values from cooperative game theory — a concept Lloyd Shapley introduced in 1953 for fairly distributing payouts among coalition players. In ML, the "players" are features, the "game" is the model's prediction, and the "payout" is the difference between the prediction and the baseline.

How SHAP works

For a model with features $x_1, x_2, ..., x_n$ producing prediction $f(x)$:

$$f(x) = \phi_0 + \sum_{i=1}^{n} \phi_i$$

Where $\phi_0$ is the baseline (average prediction) and $\phi_i$ is the SHAP value for feature $i$ — its contribution to moving the prediction away from the baseline.

Key properties:

Efficiency: SHAP values sum to the difference between the prediction and the baseline
Symmetry: Features with equal impact get equal SHAP values
Additivity: SHAP values from ensemble models decompose into per-model contributions

TreeExplainer vs. KernelExplainer

Explainer	Models Supported	Speed	Exact?	When to Use
TreeExplainer	Decision trees, random forests, gradient boosting	Fast	Yes (exact Shapley values)	Always prefer for tree-based models
KernelExplainer	Any model	Slow (O(2^n) approximations)	Approximate	Use for non-tree models (SVM, neural nets)
DeepExplainer	Neural networks (Keras/TensorFlow)	Medium	Approximate	Use for deep learning models
GradientExplainer	Neural networks	Medium	Approximate	Alternative to DeepExplainer

KernelExplainer performance
KernelExplainer is model-agnostic but computationally expensive. On a dataset with 47 features and 1000 samples, it can take 10+ minutes. Always subsample to 100-200 rows for KernelExplainer. For tree-based models, TreeExplainer is 100x faster and produces exact values — use it whenever possible.

Unlock the full lesson

You've read the first 2 sections. The rest of this lesson covers The Concept, Check Yourself, Key Terms & Next — plus a hands-on lab, quiz, and project artifact.

Create a free account — the opening phases of 29 of 35 courses are free, no credit card.