Interpreting PDPs
A PDP conveys the marginal effect of a feature on the prediction throughout all (or interpolated) possible values for that feature. It's a global model interpretation method that can visually demonstrate the impact of a feature and the nature of the relationship with the target (linear, exponential, monotonic, and so on).
It can also be extended to include two features, to illustrate the effect of their interaction on the model. One feature plot shows in the y axis the predicted outcome or relative change in this outcome, and the x axis shows all possible values of the feature. The plotted line is calculated by changing the value of the feature to the one in the x axis for all the observations and averaging the predictions if this single feature were to change, to get the y axis coordinate.
One variation of the PDP deducts the expected value for all observations from the y axis, thus centering the marginal effect to the expected value. Another PDP variation...
