Accumulated Local Effects
ALE plots show how a feature affects a model's predictions by looking only at small local changes among genuinely similar observations, which avoids the misleading extrapolation that partial dependence plots produce when features are correlated.
A partial dependence plot estimates a feature's effect by holding every other feature fixed at its observed value and sweeping the feature of interest across its whole range — but when features are correlated, this creates fake, never-seen combinations. Sweeping "square footage" up for a house while holding "number of bedrooms" fixed at 1 eventually produces a 5,000 square-foot one-bedroom house that never occurs in real data, and the model's prediction for it is pure extrapolation, not a real pattern.
Accumulated Local Effects plots avoid the fake-combination problem of partial dependence by only comparing predictions across small local windows of genuinely observed, correlated feature combinations, then summing up those local differences — giving a much more trustworthy picture of a feature's effect when features are correlated with each other.
Instead of sweeping the feature across its full range against fixed values of everything else, ALE looks within narrow windows of the feature's actual observed values, computes how the prediction changes for the real, nearby data points in that window as the feature moves slightly, and accumulates these small, locally realistic changes into an overall curve.
Worked example. A house-price model includes both square footage and number of bedrooms, which are naturally correlated. A partial dependence plot might show an unrealistic dip at very large square footage because it's forced to average in nonsensical combinations like a huge house with one bedroom. The ALE plot for square footage instead only compares houses of, say, 1,900 vs. 2,000 square feet among homes that actually exist near that range, correctly showing a smooth, monotonic increase in predicted price without the artifact.
ALE plots are generally preferred over partial dependence plots whenever the features being examined are correlated, which in financial datasets — where balance-sheet ratios routinely move together — is closer to the rule than the exception.
Further reading
- Apley and Zhu, 'Visualizing the Effects of Predictor Variables in Black Box Supervised Learning Models'