Partial Dependence Plots
Feature importance says which inputs matter, but not how they shape the output. A partial dependence plot shows the average effect of one feature on the prediction.
How it is built
To find the partial dependence at one value of a feature, the method fixes that feature to the value for every row, keeps all other features as they are, predicts, and averages the results. Repeating across a grid of values traces a curve.
- The curve reveals shape, whether the effect is rising, flat, or curved.
- It marginalizes over the other features by averaging them out.
- The y axis is the average prediction, not a single example.
What it reveals
A partial dependence plot can expose nonlinearities and thresholds a model learned, such as risk that climbs sharply past a certain age. This makes an otherwise opaque model easier to explain.
The independence caveat
The averaging assumes the chosen feature is roughly independent of the others. When features are strongly correlated, fixing one while shuffling realistic combinations creates impossible data points, and the curve can mislead. Individual conditional expectation curves help spot when averaging hides differences across individuals.
Key idea
A partial dependence plot averages predictions over the data to show how one feature moves the output.