Chapter 1.11 — Explainability, interpretability and fairness

🎯 Objective

Cover SHAP, LIME, PDP and the limits of these techniques — as well as fairness.

🧠 Concepts

• Global interpretability — explanations that hold for the whole model.
• Local interpretability — explanations for a specific prediction.
• SHAP — approximate Shapley values, based on game theory.
• LIME — a local linear model approximating the behavior around an instance.
• PDP (Partial Dependence Plot) — the average effect of a feature.

⚠️ Limits

• SHAP and LIME are approximations — they are not "the truth" of the model.
• They can be unstable (same model, different explanations).
• They do not replace human auditing in critical decisions.

🧠 Fairness

Common metrics:

• Demographic parity — equal positive rate across groups.
• Equalized odds — equal TPR and FPR across groups.
• Calibration parity — equal calibration across groups.

Trade-off: not all metrics can be satisfied simultaneously. The choice is normative, not technical.

📌 Checklist

• [ ] Are per-segment metrics reported?
• [ ] Is fairness documented in the model card?
• [ ] Is there human review for critical decisions?

📚 References

• SHAP docs: https://shap.readthedocs.io/
• Ribeiro, Singh & Guestrin — "Why Should I Trust You?" Explaining the Predictions of Any Classifier (LIME, KDD 2016): https://arxiv.org/abs/1602.04938
• Hardt, Price & Srebro — Equality of Opportunity in Supervised Learning (NIPS 2016): https://arxiv.org/abs/1610.02413
• Barocas, Hardt & Narayanan — Fairness and Machine Learning: Limitations and Opportunities (open book): https://fairmlbook.org/

---