Explainable AI and Model Interpretability
SHAP, LIME, feature importance and explaining decisions to stakeholders.
Taught by Bibek Sharma · Data Scientist
What you'll learn
- Distinguish global from local explanations and choose the right method for a question
- Fit and interpret inherently interpretable models including generalised additive models
- Compute permutation importance and partial dependence plots and read them correctly
- Explain individual predictions with LIME and SHAP and check their stability
- Audit a model for disparate impact and other fairness issues across protected groups
- Communicate model behaviour to stakeholders with honest, well-scoped explanations
About this course
This explainable AI course is for data scientists, risk analysts and product owners who must explain what a machine learning model does and why, whether to a regulator, a customer or a sceptical executive. Credit scoring, insurance, healthcare and hiring all demand explanations, and interpretability is now a core skill rather than an afterthought. You will learn the difference between global and local explanations, use permutation importance and partial dependence to understand a model overall, apply LIME and SHAP to explain individual predictions, and run fairness checks that surface disparate impact across groups. You will also learn when an interpretable model such as a generalised additive model is the better choice, and how to present explanations to non-technical stakeholders without overclaiming. You should be comfortable with Python and scikit-learn. By the end you will be able to audit a model, defend its decisions and document them.
Course content
7 lessons- 1Why Explanations MatterThe case for interpretability, the global versus local distinction, and the running credit-risk example.Preview
- 2Interpretable Models by DesignLinear models, small trees and generalised additive models that are transparent without sacrificing much accuracy.
- 3Global Methods: Permutation Importance and Partial DependenceMeasure what the model relies on and how predictions move with each feature, and avoid the classic misreadings.
- 4Local Explanations With LIMEExplain one prediction by fitting a simple model around it, and learn how to judge whether the explanation is stable.
- 5SHAP in PracticeShapley values for consistent local and global explanations, with TreeExplainer, summary plots and dependence plots.
- 6Fairness Checks and Bias AuditsMeasure disparate impact and error-rate gaps across groups, find proxy features, and decide how to respond.
- 7Communicating Models to StakeholdersTranslate explanations into decisions, avoid overclaiming, and document a model so it survives review.