shapinteractions 0.1

Build a comprehensive interaction graph visualization

README

This package allows you to create a comprehensive visualization graph of feature interactions for machine learning models explained with Local Additive Explanation methods. It should work with any 3-dimensional array of explanation interactions but is particularly well suited for the output of the TreeExplainer SHAP interaction extractor from the shap Python library.

Get started

To get started, simply run:

pip install shapinteractions

You should be ready to go. You can now import ShapInteractions from shapinteractions and use it in your Python environment and scripts:

from shapinteractions import ShapInteractions

How to use the package?

Please head over to our GitHub repository for an example notebook.

How to read the graph?

You can interact with our interaction graph example right here.

Each feature is represented by a node:

• its color informs about whether the feature is positively correlated (red) or negatively correlated (blue) with the model predictions
• its size relates to the average absolute SHAP value of the feature's main effect (the contribution of the feature alone, without accounting for its interactions)

Each interaction is represented by an arrow:

• its color informs about whether the interaction reinforces (red) or attenuates (blue) the main effect of the pointed feature (regardless of that feature's color)
• its width relates to the average absolute SHAP interaction value
• the top slider allows you to hide/reveal interactions based on their strength (average absolute SHAP interaction value)

A node/arrow is rendered black if its Spearman's and Pearson's coefficients are of opposite signs, suggesting that the relationship is more complex than it appears. Similarly, if Spearman's coefficient for a given interaction is smaller than spearmans_threshold (default value of 0.3), the arrow is rendered dashed, suggesting that the correlation is weak.

You can hover over a node/arrow to display the corresponding correlation coefficients and average absolute SHAP value.