Despite of recent advances, existing GNN explainers usually suffer from one or more of the following issues:
-
Post-hoc explanation: Most explainers are post-hoc, in which another interpretive model needs to be created to explain a well-trained GNN.
-
Ignorance of causal-effect relationships: Most GNN explainers recognize predictive subgraphs only by the input-outcome associations rather than their intrinsic causal relationships, which may lead to the obtained explanations contain spurious correlations that are not trustable.
-
Small-scale evaluations: In biomedical fields such as bioinformatics and neuroimaging, most GNN explainers are just applied to small-scale datasets, such as molecules.
Thus, we propose a new built-in interpretable GNN to adress these issues. Our developed CI-GNN enjoys a few unique properties: the ability to produce instance-level explanation on edges; the causal-driven mechanism; and the ability to learn disentangled latent representations
Figure 1 Architecture of our proposed CI-GNN. The model consists of four modules: GraphVAE, causal effect estimator, causal subgraph generator and a basic classifier
Note that we require 1.7.0 <= torch_geometric <= 2.0.2
.
Simply run python main.py
to reproduce the results of MUTAG in the paper.
If you have any questions, suggestions, or would like to collaborate us on relevant topics, please feel free to contact us by yusj9011@gmail.com (Shujian Yu), kzzheng@stu.xjtu.edu.cn (Kaizhong Zheng).
@article{zheng2024ci,
title={Ci-gnn: A granger causality-inspired graph neural network for interpretable brain network-based psychiatric diagnosis},
author={Zheng, Kaizhong and Yu, Shujian and Chen, Badong},
journal={Neural Networks},
pages={106147},
year={2024},
publisher={Elsevier}
}