Advanced level · Module 1: Ensemble Methods · Lesson 4 of 4
Interpreting Ensembles
Find out which features drive an ensemble's predictions and how, using model-agnostic methods.
In this lesson you will
- Compute permutation importance
- Read partial dependence plots
Key concepts
permutation importanceimpurity importance pitfallspartial dependencecorrelated features
The lesson
Permutation importance shuffles one feature at a time on validation data and measures how much performance drops. A large drop means the model relies on that feature.
Built-in impurity importances are fast but biased towards features with many distinct values. Correlated features share importance, so a feature may look unimportant because a related one carries the signal.
Partial dependence plots show how the average prediction changes as one feature changes — for example food-insecurity risk rising sharply when the distance to market exceeds a certain value. They show model behaviour, not causes.
Step by step
- Compute permutation importance on validation data.
- Compare it with impurity importance.
- Plot partial dependence for the top two features.
- Write cautious interpretations.
Practice activity
Explain the food-insecurity model's behaviour with permutation importance and partial dependence.
Expected result: Correct plots and interpretations that describe model behaviour without claiming causation.
Check your understanding
-
How does permutation importance work?
Show answer
It measures the drop in performance when a feature's values are shuffled.
-
Do partial dependence plots show causes?
Show answer
No; they show how the model's predictions respond to a feature.
Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.