This repository implements a large number of removal-based explanations, a class of model explanation approaches that unifies many existing methods (e.g., SHAP, LIME, Meaningful Perturbations, L2X, permutation tests). Our paper presents a framework that allows us to implement many of these methods in a lightweight, modular codebase.

Our implementation does not take advantage of certain approximation approaches that make these methods fast in practice, so you may prefer to continue using the original implementations (e.g., SHAP, LIME, SAGE). We also haven't implemented every method, e.g., we do not support image blurring or feature selection approaches.

To begin, you need to clone the repository and install the library into your Python environment:

pip install .

Our code is designed around the framework described in the paper. Each model explanation method is specified by three choices:

1. Feature removal: how the model is evaluated when features are held out
2. Model behavior: a target quantity that's analyzed as features are removed (e.g., an individual prediction, or the model loss)
3. Summary technique: how each feature's influence is summarized (e.g., using Shapley values)

The general use pattern looks like this:

from rexplain import removal, behavior, summary

# Get model and data
x, y = ...
model = ...

# 1) Feature removal
extension = removal.MarginalExtension(x[:512], model)

# 2) Model behavior
game = behavior.PredictionGame(x[0], extension)

# 3) Summary technique
attr = summary.ShapleyValue(game)
plt.bar(np.arange(len(attr)), attr)

For usage examples, see the following notebooks:

• Census shows how to explain individual predictions
• MNIST shows how to explain the model's loss for individual predictions
• Breast cancer (BRCA) shows how to explain the dataset loss

• Ian Covert ([email protected])
• Scott Lundberg
• Su-In Lee

Ian Covert, Scott Lundberg, Su-In Lee. "Explaining by Removing: A Unified Framework For Model Explanation." arXiv preprint:2011.14878