Accumulated Local Effects

Unlike Partial Dependence Plots (PDP), ALE does not rely on the marginal distribution of features. It computes effects by conditioning on local neighborhoods, isolating the effect of the target feature even when it is highly correlated with others. This prevents the extrapolation of predictions into unrealistic regions of the feature space, a critical flaw of PDPs.

The core algorithm partitions the feature into intervals. Within each interval, it calculates the first-order difference in prediction by replacing the feature value with the upper and lower bounds of the interval. These local effects are then accumulated (integrated) across intervals and centered to produce the final ALE curve, isolating the main effect.

The raw accumulated effect is uncentered and lacks a meaningful zero point. ALE subtracts the average effect across all instances to produce a centered curve, where the mean effect is zero. This allows direct interpretation: the ALE value at a feature level represents the difference from the average prediction.

ALE extends naturally to visualize two-way interactions. By computing the joint effect of two features and subtracting their individual main effects, ALE isolates the pure interaction term. This reveals synergistic or antagonistic relationships that are invisible in univariate plots.

ALE is significantly faster than model-agnostic Shapley value methods. It requires only K passes over the data, where K is the number of intervals, making it scalable to large datasets. This efficiency makes it suitable for rapid exploratory analysis during model development.

ALE plots show the average marginal effect of a feature. A flat line indicates no effect. A positive slope means increasing the feature increases the prediction. Crucially, ALE assumes feature independence within each local interval; very high local correlations can still introduce bias.

A disaggregated view of PDPs. Instead of averaging, ICE plots the prediction curve for every single instance. This reveals heterogeneous effects that PDPs hide. While ICE inherits PDP's extrapolation problem under correlation, combining ICE with ALE provides both instance-level granularity and unbiased aggregated trends.

A model-agnostic global method that measures feature importance by the drop in model performance after randomly shuffling a feature column. Unlike ALE, it does not reveal the functional form or direction of the effect. It is also susceptible to inflating importance for correlated features, a bias ALE's conditional partitioning avoids.

A refinement of ICE plots that anchors all curves to zero at a specific feature value, making it easier to visualize heterogeneous effects. While c-ICE improves visual comparison, it remains a PDP derivative and still suffers from the extrapolation bias in correlated regions. ALE remains the preferred unbiased alternative for the main effect curve.