Game Theory and XAI

Welcome to the documentation for the myerson package. 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 and Chemprop models 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, for example:
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

For example code have a look on the Get Started page.

Contents

• Home
• Get started
  • A Brief Introduction to Cooperative Game Theory
    • The Gloves Game
  • Code Examples
    • Calculate the Myerson Values
    • Explain a PyG GNN Prediction
    • Visualizing Explanations on Molecular Structures
• Documentation
  • Calculate the Myerson value for arbitrary games
    • MyersonCalculator
    • MyersonSampler
  • Explain PyG GNNs with Myerson values
    • MyersonExplainer
    • MyersonSamplingExplainer
    • MyersonClassExplainer
    • MyersonSamplingClassExplainer
  • Explain Chemprop Models with Myerson values
• GitHub ⧉
• ChemRxiv ⧉