Paradigm Shift of GNN Explainer from Label Space to Prototypical Representation Space. ICLR 2026, Poster.

---

Environment

The environment has been exported to IDEA/requirements.txt.

Dataset

Regrading the four real-wolrd datasets, i.e., Mutag, Alkane, Fluoride, Benzene, the original resource is available at Harvard Dataverse, which is released by Evaluating explainability for graph neural networks.

Furthermore, the processed datasets alongside the ground explanantions are uploaled in IDEA/data/.

Model

• The architecture of the target GNN to be explained in defined in IDEA/gnn_model.py. In IDEA/GNN, we upload the weights of trained GNN models.
• The RQ-VAE module is implemented in IDEA/RQ-VAE.py. Afterwards, the hierarchical graph tokenizer (HGTokenizer) is implemented in IDEA/graph_tokenizer.py.
• The IDEA explainer with PGExplainer as backbone is implemented in IDEA/explainer/idea.py. In addition, we upload the weights of pre-trained codebooks, including both shallow and deep branches, in IDEA/codebook.

Evaluation

Based on IDEA/explainer/eval_idea.py, one can optimize and evaluate the IDEA explainer. The launching command belike,

python IDEA/explainer/eval_idea.py --dataset Benzene --root_path IDEA/ --epochs 10 --lr 0.0005 --device -1 --batch_size 64 --codebook_size 32 --gt_learning_rate 0.01 --gt_epoch 10

---

@inproceedings{idea2026,
  title     = {{Paradigm Shift of GNN Explainer from Label Space to Prototypical Representation Space}},
  author    = {Yin, Jun and Wang, Senzhang and Luo, Ziluowen and Huo, Peng and Yan, Hao and Miao, Hao and Li, Chaozhuo and Pan, Shirui and Zhang, Chengqi},
  booktitle = {International Conference on Learning Representations},
  year      = {2026},
  url       = {https://openreview.net/forum?id=X7eYISNf01}
}