5.20 Model interpretability
You can answer 'why did it say that' for any model you ship.
Before:03. Data Handling & AnalysisUnlocks:06. Deep Learning10. Production & MLOps13. Capstones, Portfolio & Interviews
Interpretability answers the question every stakeholder eventually asks — why did it say that — with permutation importance and partial dependence for global behaviour, SHAP for a single prediction. In regulated and high-stakes settings this is a shipping requirement, not a nicety. It sits at the end of the classical module because it examines whatever got built. The boundary to hold: explanations attribute the model's behaviour, not the world's causes, and reading SHAP causally repeats the correlation mistake with better graphics.
Work through these
Global vs. local explanation
Explaining a model's overall behaviour and explaining one particular prediction are different questions with different methods. Confusing them leads to answers that do not satisfy the person asking.
Permutation importance and its pitfalls
Shuffling a feature and measuring the damage gives an importance estimate that works for any model, and it misleads badly when features are correlated. That caveat is essential rather than optional.
SHAP values and force plots
A method with a principled basis for attributing a prediction among the features, along with its standard visualisations. It has become the default expectation in many workplaces.
Partial dependence and ICE curves
Plots showing how the prediction changes as one feature varies, either averaged or per individual. The individual version reveals variation the average hides.
Sign in to keep your progress.
Free resources
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.