GEM-FIX

Unlocking the Game: Estimating Essential Games in Mobius Representation for Explanation and High-order Interaction Detection

Introduction

This package inlcudes the implementation of GEM-FIX: Game Estimation in Mobius representation for Feature Interaction detection eXplanation

Features

• Model Agnostic: GEM-FIX can explain any model
• Shapley Value: GEM-FIX is a additive feature attribution method based on the Shapley value
• Interaction Detection GEM-FIX is able to detect high-order significant interactions from exponentially many possibilities

Credits

GEM-FIX is used the SHAP (See SHAP) infrastructure for the implementaiton

Requirements

See requirements.txt for all the requirements.

Usage

Below is an example illustrating how to use GEM-FIX:

import transformers
from explainer.gemfix import GEMFIX
 

# load a transformers pipeline model
model = transformers.pipeline('sentiment-analysis', return_all_scores=True)

# explain the model on a sample
explainer = gemfix.Explainer(model)
gemfix_values = explainer.shap_values(["What a great movie! ...if you have no taste."])