hoocs 0.0.1

Occlusion-based explainers for higher-order attributions.

Higher Order OCclusion (hoocs)

Introduction

Hoocs implements a broad range of model-agnostic attributions

• PredDiff [1] [2]
• Shapley values [3]
• KernelSHAP [4]

Recently, there has been increasing interest in more in-depth analysis of models. To meet this needs, the analysis of feature interactions is inevitable. Therefore, this package allows to calculate arbitrary higher-order explanations. Tt is extendable to other methods, which rely on marginalizing features in input space.

Installation

pip install hoocs

Implement new imputers

To enable reliable attributions, hooks enables simple incorporation of custom imputers for any kind of data modality. To add a new imputer, the user is requested to inherent from the abstract base Imputer class in hoocs.imputers.abstract_imputers.py. This class performs basic type checking and ensures a consistent interface.

References

[1] Explaining classifications for individual instances.

[2] PredDiff: Explanations and interactions from conditional expectations

[3] An efficient explanation of individual classifications using game theory

[4] A Unified Approach to Interpreting Model Predictions