Calculate Myerson values and explain GNNs

This package implements the Myerson solution concept from cooperative game theory. The Myerson values attribute every player of a game their fair contribution to the games payoff. Myerson values are related to Shapley values but the player cooperation is restricted by a graph.

A graph neural network (GNN) can be treated as a coalition function for a game and the Myerson values can be used as feature attribution explanations to understand a model prediction. This package also implements Methods to explain PyG GNNs with Myerson values.

Calculating the Myerson value scales exponentially with bigger graphs / more players. Therfore, Monte Carlo sampling techniques were implemented to approximate the Myerson values.

Installation

Install the complete package with PyTorch dependencies using one of the following commands:

# pip
pip install myerson[explain]

# conda / mamba
conda install myerson

# for conda / mamba, you need to manually install pytorch dependencies
conda install pytorch torchvision torchaudio cpuonly -c pytorch
conda install conda install pyg -c pyg

If you are only interested in the game theoretic part you don't need to install PyTorch:

# pip 
pip install myerson

# conda / mamba
conda install myerson

Examples and Documentation

Example uses can be found here. The full documentation can be found at https://myerson.readthedocs.io/.

Results of Explaining GNNs for Cheminformatics

This package was developed to explain GNN predicitons for Cheminformatics approaches. The repository with code and data for these experiments can be found here: https://github.com/kochgroup/myerson_results.

Citation

TBD.