Deep Dive on Accumulated Local Effect Plots (ALEs) with Python

Intuition, algorithm and code for using ALEs to explain machine learning models

(source: author)

Highly correlated features can wreak havoc on your model interpretations. They violate the assumptions of many XAI methods and make it difficult to understand the nature of a feature’s relationship with the target. At the same time, it is not always possible to remove them without affecting performance. We need a method that can provide clear interpretations, even with multicollinearity. Thankfully we can rely on ALEs [1].

ALEs are a global interpretation method. Like PDPs they show the trends captured by the model. That is if a feature has a linear, non-linear or no relationship with the target variable. However, we will see that the method of identifying these trends is quite different. We will:

• Give you the intuition for how ALEs are created.
• Formally define the algorithm used to create ALEs.
• Apply ALEs using the Alibi Explain package.

We will see that, unlike other XAI methods like SHAP, LIME, ICE Plots and Friedman’s H-stat, ALEs give interpretations that are robust to multicollinearity.

You may also enjoy this video on the topic. And, if you want to learn more, check out my course — XAI with Python. You can get free access if you sign up to my newsletter.

Understanding ALEs

We will use the abalone dataset to understand how ALEs work. Abalone is a shellfish delicacy. We want to predict the number of rings in their shell using features like shell weight and shucked weight (weight of the meat). Figure 1 shows the correlation heatmap for all the numerical features in this dataset. We are dealing with some highly correlated features!