Classical ML, validation, metrics, interpretation, debugging · insight-0186

Switch from PDP to ALE when the feature of interest is strongly correlated with others

Situation
You need the effect shape of a feature that moves with others, such as aircraft capacity and route distance, or price and season.
Action
Use accumulated local effects, which average prediction differences within small intervals of the feature using only nearby observed rows, and show the data density along the axis.
Reason
ALE conditions on the feature's local neighbourhood, so it avoids extrapolating into regions without data while still isolating the feature's effect.
Boundary
Black-box models with correlated inputs where a global effect plot is needed.

Symptom or risk

The PDP evaluates the model on impossible combinations (a wide-body on a 300 km route) and the curve reflects extrapolation, not data.

With roughly independent features, PDP and ALE agree and the PDP is easier to explain.

ALE is less familiar to stakeholders, depends on interval choice and has no per-instance view like ICE.

Reading any effect plot as causal; ALE describes the model's behaviour, not the outcome of an intervention.

ALE plots avoid the extrapolation problem of PD plots under correlated predictors and are cheaper to compute.
Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models
PDP and ICE assume the features of interest are independent of the others and can create unrealistic data points when they are correlated.
Partial Dependence and Individual Conditional Expectation plots (scikit-learn User Guide 5.1)