dmitrykazhdan / gci

GCI: A Graph Concept Interpretation Framework

GCI: A Graph Concept Interpretation Framework

This repository contains an implementation of GCI. GCI is a (G)raph (C)oncept (I)nterpretation Framework, which can be used for quantitatively verifying concepts extracted from GNNs, using provided human interpretations.

Specifically, GCI encodes user-provided concept interpretations as functions, which can be used to quantitatively measure alignment between extracted concepts and user interpretations.

Abstract

Explainable AI (XAI) underwent a recent surge in research on concept extraction, focusing on extracting human-interpretable concepts from Deep Neural Networks. An important challenge facing concept extraction approaches is the difficulty of interpreting and evaluating discovered concepts, especially for complex tasks such as molecular property prediction. We address this challenge by presenting GCI: a (G)raph (C)oncept (I)nterpretation framework, used for quantitatively measuring alignment between concepts discovered from Graph Neural Networks (GNNs) and their corresponding human interpretations. GCI encodes concept interpretations as functions, which can be used to quantitatively measure the alignment between a given interpretation and concept definition. We demonstrate four applications of GCI: (i) quantitatively evaluating concept extractors, (ii) measuring alignment between concept extractors and human interpretations, (iii) measuring the completeness of interpretations with respect to an end task and (iv) a practical application of GCI to molecular property prediction, in which we demonstrate how to use chemical functional groups to explain GNNs trained on molecular property prediction tasks, and implement interpretations with a 0.76 AUCROC completeness score.

Citing

If you find this code useful in your research, please consider citing:

@article{kazhdan2023gci,
  title={GCI: A (G) raph (C) oncept (I) nterpretation Framework},
  author={Kazhdan, Dmitry and Dimanov, Botty and Magister, Lucie Charlotte and Barbiero, Pietro and Jamnik, Mateja and Lio, Pietro},
  journal={arXiv preprint arXiv:2302.04899},
  year={2023}
}