Pith. sign in

REVIEW 38 references

Towards Comprehensive and Prerequisite-Free Explainer for Graph Neural Networks

T0 review · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read OPEN clusters training graphs into inferred environments and trains a variational subgraph generator to explain GNN predictions across distribution shifts without model internals or edge weights.

arxiv 2505.14005 v2 pith:MHEDUXLT submitted 2025-05-20 cs.LG cs.AI

Han Zhang , Yan Wang , Guanfeng Liu , Pengfei Ding , Huaxiong Wang , Kwok-Yan Lam This is my paper · ORCID
classification cs.LGcs.AI
keywords gnnsdecisionlogicopentextbfacrosscompletedataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Graph neural networks (GNNs) make predictions by passing messages between connected nodes. To explain a prediction, most methods search for a small subgraph that, on its own, makes the GNN give the same answer. OPEN is a new version of this idea that tries to work across distribution shifts: the model is trained on one kind of graph, but tested on graphs that look different.

OPEN has two main parts. The first part, NPAF, turns each graph into a structural embedding using Weisfeiler-Leman features and clusters the training graphs with K-means. Each cluster is treated as an environment with its own distribution. It then tries to separate nodes and feature dimensions that depend on the environment from those that depend on the label. The second part, GVAG, is a variational generator. For each graph it produces a probability for every node and edge, and samples explanation subgraphs from those probabilities. A set of losses encourages the sampled subgraph to trigger the same GNN prediction as the original graph, while contrastive and reward terms try to keep the generator stable across environments.

The paper reports that OPEN beats existing explainers on fidelity on the GOOD Cora and Motif benchmarks, and that it can run without access to GNN internals or learnable edge weights. The main limitation is that fidelity is measured as whether the subgraph reproduces the GNN label, which is exactly what OPEN is trained to do, so it does not prove that OPEN found the model's true reasoning.

Extended reading notes

Core claim

The abstract states: OPEN "can infer and partition the entire dataset's sample space into multiple environments" and "captures nearly complete decision logic of GNNs ... outperforms state-of-the-art methods in fidelity while maintaining similar efficiency, and enhances robustness in real-world scenarios." If correct, OPEN is the first GNN explainer that works both without GNN internal access and without learnable edge weights, while providing faithful explanations under covariate and concept distribution shifts.

Load-bearing premise

The load-bearing premise is that K-means clustering on Weisfeiler-Leman structure embeddings and JS-divergence feature analyses correctly recover the true underlying environments E, and that graphs decompose cleanly into a label-relevant part Gc and an environment-relevant part Gs. This enters in Section 4.1 ("Obtain Structure-Based Embedding", "Infer Potential Environmental Label Based on Structure") and is directly contradicted for the Motif basis domain, where the paper admits in Section 5.1 that the two-level OOD issue "does not align with the SCM used by OPEN, which limits NPAF's ability to partition the sample space."

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

OPEN rests on a clean Gc/Gs causal split and on K-means recovering true environments from structural embeddings and feature-dimension distributions. Neither assumption is externally validated, and the Motif basis results show a concrete failure mode when the clean split does not hold.

free parameters (4)
  • Number of environments K = 4 (Cora), 5 (Motif)
    Chosen per dataset in Table 7. K directly controls the granularity of the environment partition and therefore the output of NPAF and all downstream embeddings.
  • Prior subgraph density = 0.35 (Cora), 0.1 (Motif)
    Selected by tuning in Table 7. It controls node and edge counts of explanations and strongly affects the density and fidelity metrics.
  • Thresholds for node-environment relevance = not specified
    Section 4.1 splits nodes into Vc and Gs based on 'high' and 'low' gradients of embedding variance, but no criterion or numeric threshold is given.
  • Loss weights (omega_RECON, omega_CON, omega_LAR) = 2, 0.5, 1
    Balance the losses in Eq. 14; they are tuned per dataset and the appendix shows they materially change the metrics.
assumptions (4)
  • domain assumption Each graph decomposes into independent components Gc (label-relevant) and Gs (environment-relevant).
    In Section 4.1 and Fig. 1, OPEN assumes this causal split. The paper admits in Section 5.1 that the Motif basis domain has a two-level OOD structure that does not follow this SCM.
  • domain assumption Graphs affected by the same environment share similar structural patterns recoverable by WL embeddings and K-means.
    Stated in Section 4.1 under 'Infer Potential Environmental Label Based on Structure'. If false, the inferred environments are arbitrary clusters.
  • domain assumption Feature dimensions influenced by the environment have consistent distributions across node types and labels, making JS-divergence isolation valid.
    Used in Section 4.1 under 'Identify Causal Dimension in Node Features' to define Dimenv.
  • domain assumption The target GNN M is causally affected by G and Y, opening a backdoor path from the explainer to Y.
    Postulated in Section 3 (Preliminaries) and used to justify learning Gc-Y relationships. It is presented without direct validation.
invented entities (1)
  • Inferred environment labels Estr and Efeat
    purpose: Latent partition of the sample space used to train invariant embeddings and to sample subgraphs per distribution.
    No ground-truth environment labels are checked. The only evidence for these environments is internal clustering consistency and downstream fidelity on a metric that OPEN is trained to optimize.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Comprehensive and Prerequisite-Free Explainer for Graph Neural Networks." pith.science (2026). https://pith.science/paper/MHEDUXLT

@misc{pith2026250514005,
  author       = {Pith},
  title        = {Pith review of: Towards Comprehensive and Prerequisite-Free Explainer for Graph Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHEDUXLT}},
  note         = {Machine review of arXiv:2505.14005}
}
read the original abstract

To enhance the reliability and credibility of graph neural networks (GNNs) and improve the transparency of their decision logic, a new field of explainability of GNNs (XGNN) has emerged. However, two major limitations severely degrade the performance and hinder the generalizability of existing XGNN methods: they (a) fail to capture the complete decision logic of GNNs across diverse distributions in the entire dataset's sample space, and (b) impose strict prerequisites on edge properties and GNN internal accessibility. To address these limitations, we propose OPEN, a novel c\textbf{O}mprehensive and \textbf{P}rerequisite-free \textbf{E}xplainer for G\textbf{N}Ns. OPEN, as the first work in the literature, can infer and partition the entire dataset's sample space into multiple environments, each containing graphs that follow a distinct distribution. OPEN further learns the decision logic of GNNs across different distributions by sampling subgraphs from each environment and analyzing their predictions, thus eliminating the need for strict prerequisites. Experimental results demonstrate that OPEN captures nearly complete decision logic of GNNs, outperforms state-of-the-art methods in fidelity while maintaining similar efficiency, and enhances robustness in real-world scenarios.

Figures

Figures reproduced from arXiv: 2505.14005 by the authors.

Figure 1
Figure 1. SCMs, where grey and white nodes indicate observable [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overview of the proposed OPEN framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Hyper-parameter sensitivity study on different weights of [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Hyper-parameter sensitivity study on different edge densi [PITH_FULL_IMAGE:figures/full_fig_p012_3.png]
Figure 5
Figure 5. Figure 5: Hyper-parameter sensitivity study on different weights of [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Explanation subgraphs [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 37 canonical work pages

  1. [1]

    Evaluating explainability for graph neural networks.Scientific Data, 10(1):144,

    [Agarwalet al., 2023 ] Chirag Agarwal, Owen Queen, Himabindu Lakkaraju, and Marinka Zitnik. Evaluating explainability for graph neural networks.Scientific Data, 10(1):144,

    2023

  2. [2]

    the model to preserve more crucial structures within the explanation subgraphs, thereby boosting positive fidelity

    This trend suggests that higher weights prompt 0.0000 0.1000 0.2000 0.3000 0.4000 0.5000 Posi�ve Fidelity Nega�ve Fidelity Unfaithfulness Node density Edge density Reocn Weight=0 Reocn Weight=0.5 Reocn Weight=1 Reocn Weight=1.5 Reocn Weight=2 Figure 5: Hyper-parameter sensitivity study on different weights of reconstruction loss. the model to preserve mor...

    work page 2000

  3. [3]

    As edge density increases from 0.1 to 1.0, a distinct trend in the performance metrics emerges: • Positive Fidelity:Variations in positive fidelity suggest that explanation subgraphs with higher densities include more critical features essential for GNN predictions, thereby enhancing positive fidelity by better aligning with GNN predictions. • Negative Fi...

    work page 2000

  4. [5]

    D4explainer: in-distribution gnn explana- tions via discrete denoising diffusion

    [Chenet al., 2023 ] Jialin Chen, Shirley Wu, Abhijit Gupta, and Rex Ying. D4explainer: in-distribution gnn explana- tions via discrete denoising diffusion. InNeurIPS, pages 78964–78986,

    work page 2023

  5. [7]

    [Fey and Lenssen, 2019] Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. InICLR Workshop,

    work page 2019

  6. [9]

    Good: A graph out-of-distribution benchmark

    [Guiet al., 2022 ] Shurui Gui, Xiner Li, Limei Wang, and Shuiwang Ji. Good: A graph out-of-distribution benchmark. InNeurIPS, volume 35, pages 2059–2073,

    work page 2022

  7. [10]

    Mean-field theory of graph neural networks in graph partitioning

    [Kawamotoet al., 2018 ] Tatsuro Kawamoto, Masashi Tsub- aki, and Tomoyuki Obuchi. Mean-field theory of graph neural networks in graph partitioning. InNeurIPS, vol- ume 31,

    work page 2018

  8. [13]

    Distribution shift detection for the postmarket surveillance of medical ai algorithms: a retro- spective simulation study.NPJ Digital Medicine, 7(1):120,

    [Kochet al., 2024 ] Lisa M Koch, Christian F Baumgartner, and Philipp Berens. Distribution shift detection for the postmarket surveillance of medical ai algorithms: a retro- spective simulation study.NPJ Digital Medicine, 7(1):120,

    work page 2024

Show all 38 references
  1. [14]

    Xgexplainer: Robust evaluation-based explanation for graph neural networks

    [Kubo and Difallah, 2024] Ryoji Kubo and Djellel Difallah. Xgexplainer: Robust evaluation-based explanation for graph neural networks. InSDM, pages 64–72. SIAM,

    2024

  2. [15]

    Graph representation learning in biomedicine and healthcare.Nature Biomedical Engineering, 6(12):1353– 1369,

    [Liet al., 2022 ] Michelle M Li, Kexin Huang, and Marinka Zitnik. Graph representation learning in biomedicine and healthcare.Nature Biomedical Engineering, 6(12):1353– 1369,

    2022

  3. [16]

    DIG: A turnkey library for diving into graph deep learning research

    [Liuet al., 2021 ] Meng Liu, Youzhi Luo, Limei Wang, Yaochen Xie, Hao Yuan, Shurui Gui, Haiyang Yu, Zhao Xu, Jingtun Zhang, Yi Liu, Keqiang Yan, Haoran Liu, Cong Fu, Bora M Oztekin, Xuan Zhang, and Shuiwang Ji. DIG: A turnkey library for diving into graph deep learning researc...

    2021

  4. [17]

    Cf-gnnexplainer: Counterfactual explanations for graph neural networks

    [Lucicet al., 2022 ] Ana Lucic, Maartje A Ter Hoeve, Gabriele Tolomei, Maarten De Rijke, and Fabrizio Silvestri. Cf-gnnexplainer: Counterfactual explanations for graph neural networks. InAISTATS, pages 4499–4511. PMLR,

    2022

  5. [18]

    Parameterized explainer for graph neural network

    [Luoet al., 2020 ] Dongsheng Luo, Wei Cheng, Dongkuan Xu, Wenchao Yu, Bo Zong, Haifeng Chen, and Xiang Zhang. Parameterized explainer for graph neural network. InNeurIPS, volume 33, pages 19620–19631. Curran Asso- ciates, Inc.,

    2020

  6. [20]

    Graph neural collaborative filtering with medical content-aware pre-training for treatment pattern recommendation.Pattern Recognition Letters, 185:210– 217,

    [Minet al., 2024 ] Xin Min, Wei Li, Ruiqi Han, Tianlong Ji, and Weidong Xie. Graph neural collaborative filtering with medical content-aware pre-training for treatment pattern recommendation.Pattern Recognition Letters, 185:210– 217,

    2024

  7. [21]

    Pope, Soheil Kolouri, Moham- mad Rostami, Charles E

    [Popeet al., 2019 ] Phillip E. Pope, Soheil Kolouri, Moham- mad Rostami, Charles E. Martin, and Heiko Hoffmann. Explainability methods for graph convolutional neural net- works. InIEEE/CVF CVPR, pages 10772–10781,

    2019

  8. [22]

    Generating robust counterfactual witnesses for graph neural networks.arXiv preprint arXiv:2404.19519,

    [Qiuet al., 2024 ] Dazhuo Qiu, Mengying Wang, Arijit Khan, and Yinghui Wu. Generating robust counterfactual witnesses for graph neural networks.arXiv preprint arXiv:2404.19519,

    2024 arXiv

  9. [23]

    Graphvae: Towards generation of small graphs using variational autoencoders

    [Simonovsky and Komodakis, 2018] Martin Simonovsky and Nikos Komodakis. Graphvae: Towards generation of small graphs using variational autoencoders. InICANN, pages 412–422. Springer,

    2018

  10. [24]

    Wasserstein weisfeiler-lehman graph kernels

    [Togninalliet al., 2019 ] Matteo Togninalli, Elisabetta Ghisu, Felipe Llinares-López, Bastian Rieck, and Karsten Borg- wardt. Wasserstein weisfeiler-lehman graph kernels. In NeurIPS, volume 32,

    2019

  11. [25]

    Graph attention networks.stat, 1050(20):10–48550,

    [Veliˇckovi´cet al., 2018 ] Petar Veli ˇckovi´c, Guillem Cucu- rull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks.stat, 1050(20):10–48550,

    2018

  12. [26]

    Pgm-explainer: Probabilistic graphical model explanations for graph neural networks

    [Vu and Thai, 2020] Minh Vu and My T Thai. Pgm-explainer: Probabilistic graphical model explanations for graph neural networks. InNeurIPS, volume 33, pages 12225–12235,

    2020

  13. [27]

    Gn- ninterpreter: A probabilistic generative model-level expla- nation for graph neural networks

    [Wang and Shen, 2023] Xiaoqi Wang and Han Wei Shen. Gn- ninterpreter: A probabilistic generative model-level expla- nation for graph neural networks. InICLR,

    2023

  14. [28]

    Discovering invariant rationales for graph neural networks

    [Wuet al., 2021 ] Yingxin Wu, Xiang Wang, An Zhang, Xi- angnan He, and Tat-Seng Chua. Discovering invariant rationales for graph neural networks. InICLR,

    2021

  15. [29]

    Heterogeneous graph knowledge enhanced stock market prediction.AI Open, 2:168–174,

    [Xionget al., 2021 ] Kai Xiong, Xiao Ding, Li Du, Ting Liu, and Bing Qin. Heterogeneous graph knowledge enhanced stock market prediction.AI Open, 2:168–174,

    2021

  16. [30]

    Adaptive hypergraph network for trust prediction

    [Xuet al., 2024 ] Rongwei Xu, Guanfeng Liu, Yan Wang, Xuyun Zhang, Kai Zheng, and Xiaofang Zhou. Adaptive hypergraph network for trust prediction. InICDE, pages 2986–2999. IEEE,

    2024

  17. [31]

    Gnnexplainer: Generating explanations for graph neural networks

    [Yinget al., 2019 ] Zhitao Ying, Dylan Bourgeois, Jiaxuan You, Marinka Zitnik, and Jure Leskovec. Gnnexplainer: Generating explanations for graph neural networks. In NeurIPS, volume 32, pages 9244–9255. Curran Associates, Inc.,

    2019

  18. [32]

    Xgnn: Towards model-level explanations of graph neural networks

    [Yuanet al., 2020 ] Hao Yuan, Jiliang Tang, Xia Hu, and Shui- wang Ji. Xgnn: Towards model-level explanations of graph neural networks. InACM KDD, pages 430–438,

    2020

  19. [33]

    Research on graph neural network in stock market.Procedia Computer Science, 214:786–792,

    [Zhanget al., 2022 ] Wenjun Zhang, Zhensong Chen, Jianyu Miao, and Xueyong Liu. Research on graph neural network in stock market.Procedia Computer Science, 214:786–792,

    2022

  20. [34]

    Mixupexplainer: Generalizing explanations for graph neural networks with data augmentation

    [Zhanget al., 2023 ] Jiaxing Zhang, Dongsheng Luo, and Hua Wei. Mixupexplainer: Generalizing explanations for graph neural networks with data augmentation. InACM KDD, pages 3286–3296,

    2023

  21. [35]

    Domain disentanglement with interpolative data aug- mentation for dual-target cross-domain recommendation

    [Zhuet al., 2023 ] Jiajie Zhu, Yan Wang, Feng Zhu, and Zhu Sun. Domain disentanglement with interpolative data aug- mentation for dual-target cross-domain recommendation. InRecSys, pages 515–527,

    2023

  22. [36]

    These hyper-parameters are carefully selected based on prior empiri- cal results and tuning experiments to balance model fidelity and efficiency. B.2 Analysis of Hyper-Parameter Sensitivity We also test the impact of several hyper-parameters on the quality of the final generat...

    2000

  23. [2014]

    Hidden in plain sight: Subgroup shifts escape ood detection

    [Kochet al., 2022 ] Lisa M Koch, Christian M Schürch, Arthur Gretton, and Philipp Berens. Hidden in plain sight: Subgroup shifts escape ood detection. InMachine Learning Research, volume 172, pages 726–740. PMLR,

    2022

  24. [2018]

    Auto-encoding variational bayes.stat, 1050:1,

    [Kingma and Welling, 2014] Diederik P Kingma and Max Welling. Auto-encoding variational bayes.stat, 1050:1,

    2014

  25. [2019]

    Deepnote-gnn: predicting hospital readmission using clin- ical notes and patient network

    [Golmaei and Luo, 2021] Sara Nouri Golmaei and Xiao Luo. Deepnote-gnn: predicting hospital readmission using clin- ical notes and patient network. InACM-BCB, pages 1–9,

    2021

  26. [2020]

    Adversarial learning targeting deep neural network classification: A comprehensive review of defenses against attacks

    [Milleret al., 2020 ] David J Miller, Zhen Xiang, and George Kesidis. Adversarial learning targeting deep neural network classification: A comprehensive review of defenses against attacks. InIEEE, volume 108, pages 402–433. IEEE,

    2020

  27. [2021]

    Graphframex: Towards systematic evaluation of explainability methods for graph neural networks

    [Amaraet al., 2022 ] Kenza Amara, Zhitao Ying, Zitao Zhang, Zhihao Han, Yinan Shan, Ulrik Brandes, and Sebas- tian Schemm. Graphframex: Towards systematic evaluation of explainability methods for graph neural networks. In NeurIPS Workshop,

    2022

  28. [2022]

    Learning causally invariant representations for out-of-distribution generalization on graphs

    [Chenet al., 2022 ] Yongqiang Chen, Yonggang Zhang, Yatao Bian, Han Yang, MA Kaili, Binghui Xie, Tongliang Liu, Bo Han, and James Cheng. Learning causally invariant representations for out-of-distribution generalization on graphs. InNeurIPS, volume 35, pages 22131–22148,

    2022

  29. [2023]

    Invariance principle meets information bottleneck for out-of-distribution gener- alization

    [Ahujaet al., 2021 ] Kartik Ahuja, Ethan Caballero, Dinghuai Zhang, Jean-Christophe Gagnon-Audet, Yoshua Bengio, Ioannis Mitliagkas, and Irina Rish. Invariance principle meets information bottleneck for out-of-distribution gener- alization. InNeurIPS, volume 34, pages 3438–3450,

    2021

  30. [2024]

    Few-shot causal repre- sentation learning for out-of-distribution generalization on heterogeneous graphs.TKDE, 37(4):1804–1818,

    [Dinget al., 2025 ] Pengfei Ding, Yan Wang, Guanfeng Liu, Nan Wang, and Xiaofang Zhou. Few-shot causal repre- sentation learning for out-of-distribution generalization on heterogeneous graphs.TKDE, 37(4):1804–1818,

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.