pith. sign in

arxiv: 2605.14884 · v2 · pith:D7XFW5K3new · submitted 2026-05-14 · 💻 cs.LG

AIMing for Standardised Explainability Evaluation in GNNs: A Framework and Case Study on Graph Kernel Networks

Magdalena Proszewska , N. Siddharth This is my paper

Pith reviewed 2026-05-19 17:22 UTC · model grok-4.3

classification 💻 cs.LG
keywords explainability evaluationgraph neural networksgraph kernel networksAIM frameworkinherently interpretable modelsXAI for GNNs
0
0 comments X

The pith

The AIM framework evaluates GNN explainability by measuring accuracy together with instance-level and model-level explanations, enabling targeted improvements such as the xGKN model.

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

The paper presents AIM as a way to assess how well inherently interpretable Graph Neural Networks explain their own decisions. Current approaches either apply post-hoc methods to black-box models or focus on narrow aspects for interpretable ones, making broad comparisons hard. AIM combines three measures—accuracy, explanations for single predictions, and explanations of the overall model—while keeping requirements loose so the same tools can apply across different networks. When the authors apply it to Graph Kernel Networks they identify concrete shortcomings and use those observations to build xGKN, which keeps the original accuracy but scores better on the explanation measures. A reader would care because this gives a practical route to compare and upgrade understanding in graph models without adding heavy new constraints.

Core claim

AIM measures Accuracy, Instance-level explanations, and Model-level explanations to evaluate explainability in inherently interpretable GNNs. Applied to Graph Kernel Networks, the measures expose specific limitations in how those networks generate explanations. The resulting insights support construction of an updated model, xGKN, that preserves high predictive accuracy while showing clearer instance-level and model-level explanations.

What carries the argument

The AIM framework, which evaluates a model by combining its predictive accuracy with separate assessments of the explanations it produces for individual instances and for its global behavior.

If this is right

  • Graph Kernel Networks can be examined for concrete explanation weaknesses using the three AIM scores.
  • An updated model xGKN can be produced that retains the original accuracy level.
  • The new model xGKN registers higher scores on both instance-level and model-level explanation measures.
  • The same AIM pipeline can be applied to other inherently interpretable networks such as prototype networks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • AIM-style scoring could later be tested on post-hoc explanation methods for standard GNNs to see whether it produces consistent rankings across model types.
  • If the three measures prove stable, they could serve as a shared benchmark when teams compare new interpretable graph models in safety-critical settings.
  • One could check whether the xGKN changes also improve performance on downstream tasks that reward human-understandable outputs.

Load-bearing premise

That accuracy together with instance-level and model-level explanation scores together give a complete enough picture of explainability to let direct changes improve the model without extra domain rules.

What would settle it

Running AIM on the original Graph Kernel Networks and finding that the derived xGKN shows no gain in instance-level or model-level scores while accuracy stays the same, or that the three measures cannot separate models with visibly different explanation qualities.

Figures

Figures reproduced from arXiv: 2605.14884 by Magdalena Proszewska, N. Siddharth.

Figure 1
Figure 1. Figure 1: Overview of a GKN with SHAPExplainer. Top : forward pass. Bottom : explanation extraction. Learnable components in yellow. Input G (size n) is processed as a set of node-centered subgraphs G1, ..., Gn. Similarity scores fsim are computed between subgraphs and kernels H1, ..., Hm using function K. Scores aggregated using fagg and class c is predicted using predictor fpred. For explanation extraction, SHAP v… view at source ↗
Figure 2
Figure 2. Figure 2: AIM metrics measured for GIN, GAT, Prototypical Network (ProtGNN), and Graph Kernel [PITH_FULL_IMAGE:figures/full_fig_p011_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: AIM metrics measured for KerGNN and XGKN with threshold: p, p−0.1 in dark grey, and p+ 0.1 in light grey (see threshold choice discussion in Section 6). A.5 Combined results For easier readability and comparison, [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
read the original abstract

Graph Neural Networks (GNNs) have advanced significantly in handling graph-structured data, but a comprehensive framework for evaluating explainability remains lacking. Existing evaluation frameworks primarily involve post-hoc explanations, and operate in the setting where multiple methods generate a suite of explanations for a single model. This makes comparison of explanations across models difficult. Evaluation of inherently interpretable models often targets a specific aspect of interpretability relevant to the model, but remains underdeveloped in terms of generating insight across a suite of measures. We introduce AIM, a comprehensive framework that addresses these limitations by measuring Accuracy, Instance-level explanations, and Model-level explanations. AIM is formulated with minimal constraints to enhance flexibility and facilitate broad applicability. Here, we use AIM in a pipeline, extracting explanations from inherently interpretable GNNs such as graph kernel networks (GKNs) and prototype networks (PNs), evaluating these explanations with AIM, identifying their limitations and obtaining insights to their characteristics. Taking GKNs as a case study, we show how the insights obtained from AIM can be used to develop an updated model, xGKN, that maintains high accuracy while demonstrating improved explainability. Our approach aims to advance the field of Explainable AI (XAI) for GNNs, providing more robust and practical solutions for understanding and improving complex models.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces the AIM framework for standardized explainability evaluation of Graph Neural Networks, particularly inherently interpretable models. AIM assesses Accuracy, Instance-level explanations, and Model-level explanations with minimal constraints for flexibility. The authors apply AIM to Graph Kernel Networks (GKNs) and Prototype Networks (PNs) as case studies, extract explanations, identify limitations, and use the resulting insights to propose an updated xGKN model that preserves high accuracy while improving explainability.

Significance. If the empirical results hold and the three-axis evaluation proves actionable and generalizable, the work could help standardize explainability assessment for GNNs beyond post-hoc methods and provide a concrete pipeline from diagnosis to model revision. The emphasis on inherently interpretable architectures and the xGKN case study are strengths if supported by quantitative validation, baselines, and reproducibility details.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (case study): the claim that AIM yields an xGKN with 'improved explainability' while 'maintaining high accuracy' is presented without quantitative results, error bars, baseline comparisons against prior GKN variants, or statistical validation, making it impossible to assess whether the data support the central improvement claim.
  2. [§3] §3 (AIM formulation): the assertion that the three measures together provide a sufficiently complete and flexible evaluation without additional constraints or domain-specific adjustments is not accompanied by a systematic argument or ablation showing that key aspects (e.g., faithfulness under distribution shift, stability across graph sizes, or alignment with human-understandable substructures) are covered; if any fall outside the three axes the pipeline to xGKN becomes under-specified.
minor comments (2)
  1. [§3.1] Clarify the precise operational definitions and scoring procedures for 'instance-level' and 'model-level' explanations so that readers can reproduce the AIM scores on new GKN or PN architectures.
  2. [§4] Add a table or figure summarizing the AIM scores for the original GKN versus xGKN (and versus PNs) with explicit metrics and confidence intervals.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below, indicating planned revisions where appropriate to strengthen the manuscript.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (case study): the claim that AIM yields an xGKN with 'improved explainability' while 'maintaining high accuracy' is presented without quantitative results, error bars, baseline comparisons against prior GKN variants, or statistical validation, making it impossible to assess whether the data support the central improvement claim.

    Authors: We acknowledge that the presentation of results for xGKN would benefit from greater rigor. The manuscript reports accuracy values and AIM scores comparing xGKN to the original GKN and other baselines, but we agree these could be augmented. In the revision we will include error bars from multiple independent runs, explicit tabular comparisons against prior GKN variants using identical metrics, and statistical significance tests to support the claims of maintained accuracy and improved explainability. revision: yes

  2. Referee: [§3] §3 (AIM formulation): the assertion that the three measures together provide a sufficiently complete and flexible evaluation without additional constraints or domain-specific adjustments is not accompanied by a systematic argument or ablation showing that key aspects (e.g., faithfulness under distribution shift, stability across graph sizes, or alignment with human-understandable substructures) are covered; if any fall outside the three axes the pipeline to xGKN becomes under-specified.

    Authors: AIM is intentionally formulated around the three axes to balance predictive fidelity with local and global interpretability while imposing minimal constraints. We will expand the discussion in §3 to provide a clearer argument that faithfulness is captured through accuracy and instance-level fidelity measures, stability through model-level consistency checks, and alignment with human-interpretable substructures via the kernel and prototype mechanisms. Although the original submission does not contain a dedicated ablation, we will add a concise justification (with supporting references) showing how these aspects fall within the existing axes; a full ablation can be included if the editor deems it necessary. revision: partial

Circularity Check

0 steps flagged

AIM framework introduction and xGKN case study show no circular reductions

full rationale

The paper introduces AIM as a new evaluation framework measuring Accuracy, Instance-level explanations, and Model-level explanations with minimal constraints, then applies it to inherently interpretable models like GKNs and PNs to extract insights and develop an updated xGKN. No equations, derivations, or self-referential steps are present that reduce any claimed prediction, improvement, or uniqueness to fitted inputs or prior self-citations by construction. The central claims rest on the framework's novelty and its use on external case studies, making the overall derivation self-contained without load-bearing circular elements.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The paper introduces a new evaluation framework without listing explicit free parameters. It relies on the domain assumption that current explainability methods are limited in the ways stated. The AIM framework itself is an invented construct whose value depends on future validation.

axioms (1)
  • domain assumption Existing evaluation frameworks primarily involve post-hoc explanations and operate in the setting where multiple methods generate explanations for a single model.
    This premise is stated directly in the abstract as the motivation for AIM.
invented entities (2)
  • AIM framework no independent evidence
    purpose: To measure Accuracy, Instance-level explanations, and Model-level explanations with minimal constraints
    Newly proposed in the paper as the central contribution.
  • xGKN model no independent evidence
    purpose: Updated graph kernel network with improved explainability while maintaining accuracy
    Derived from insights obtained via AIM in the case study.

pith-pipeline@v0.9.0 · 5771 in / 1397 out tokens · 46506 ms · 2026-05-19T17:22:30.172434+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

67 extracted references · 67 canonical work pages

  1. [1]

    GNNExplainer: Generating Explanations for Graph Neural Networks , url =

    Zhitao Ying and Dylan Bourgeois and Jiaxuan You and Marinka Zitnik and Jure Leskovec , bibsource =. GNNExplainer: Generating Explanations for Graph Neural Networks , url =. Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada , editor =

    work page 2019

  2. [2]

    Parameterized Explainer for Graph Neural Network , url =

    Dongsheng Luo and Wei Cheng and Dongkuan Xu and Wenchao Yu and Bo Zong and Haifeng Chen and Xiang Zhang , bibsource =. Parameterized Explainer for Graph Neural Network , url =. Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , editor =

    work page 2020

  3. [3]

    KerGNNs: Interpretable Graph Neural Networks with Graph Kernels , url =

    Aosong Feng and Chenyu You and Shiqiang Wang and Leandros Tassiulas , bibsource =. KerGNNs: Interpretable Graph Neural Networks with Graph Kernels , url =. Thirty-Sixth

    work page

  4. [4]

    ProtGNN: Towards Self-Explaining Graph Neural Networks , url =

    Zaixi Zhang and Qi Liu and Hao Wang and Chengqiang Lu and Cheekong Lee , bibsource =. ProtGNN: Towards Self-Explaining Graph Neural Networks , url =. Thirty-Sixth

    work page

  5. [5]

    arXiv , author =:2112.07436 , primaryclass =

    work page arXiv

  6. [6]

    The Co-12 Recipe for Evaluating Interpretable Part-Prototype Image Classifiers , url =

    Meike Nauta and Christin Seifert , journal =. The Co-12 Recipe for Evaluating Interpretable Part-Prototype Image Classifiers , url =

    work page

  7. [7]

    and Debnath, Gargi and Shusterman, Alan J

    Debnath, Asim Kumar and Lopez de Compadre, Rosa L. and Debnath, Gargi and Shusterman, Alan J. and Hansch, Corwin , doi =. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds. Correlation with molecular orbital energies and hydrophobicity , url =. https://doi.org/10.1021/jm00106a046 , journal =

    work page doi:10.1021/jm00106a046

  8. [8]

    Pinar Yanardag and S. V. N. Vishwanathan , bibsource =. Deep Graph Kernels , url =. Proceedings of the 21th. doi:10.1145/2783258.2783417 , editor =

    work page doi:10.1145/2783258.2783417

  9. [9]

    Bioinformatics21 (06 2005)

    Borgwardt, Karsten M. and Ong, Cheng Soon and Sch\". Protein function prediction via graph kernels , url =. Bioinformatics , number =. doi:10.1093/bioinformatics/bti1007 , issn =

    work page doi:10.1093/bioinformatics/bti1007

  10. [10]

    Global Explainability of GNNs via Logic Combination of Learned Concepts , url =

    Steve Azzolin and Antonio Longa and Pietro Barbiero and Pietro Li. Global Explainability of GNNs via Logic Combination of Learned Concepts , url =. The Eleventh International Conference on Learning Representations,

    work page

  11. [11]

    Lundberg and Su

    Scott M. Lundberg and Su. A Unified Approach to Interpreting Model Predictions , url =. Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA,

    work page 2017

  12. [12]

    Graph Kernel Convolutions for Interpretable Classification , year =

    Magdalena Proszewska and Siddharth N , html =. Graph Kernel Convolutions for Interpretable Classification , year =. Data-centric Machine Learning Research (DMLR) Workshop at the International Conference on Learning Representations (ICLR) , publisher =

    work page

  13. [13]

    How Powerful are Graph Neural Networks? , url =

    Keyulu Xu and Weihua Hu and Jure Leskovec and Stefanie Jegelka , bibsource =. How Powerful are Graph Neural Networks? , url =. 7th International Conference on Learning Representations,

    work page

  14. [14]

    Explainability Techniques for Graph Convolutional Networks , url =

    Federico Baldassarre and Hossein Azizpour , journal =. Explainability Techniques for Graph Convolutional Networks , url =

    work page

  15. [15]

    Vu and My T

    Minh N. Vu and My T. Thai , bibsource =. PGM-Explainer: Probabilistic Graphical Model Explanations for Graph Neural Networks , url =. Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , editor =

    work page 2020

  16. [16]

    ter Hoeve and Gabriele Tolomei and Maarten de Rijke and Fabrizio Silvestri , bibsource =

    Ana Lucic and Maartje A. ter Hoeve and Gabriele Tolomei and Maarten de Rijke and Fabrizio Silvestri , bibsource =. CF-GNNExplainer: Counterfactual Explanations for Graph Neural Networks , url =. International Conference on Artificial Intelligence and Statistics,

    work page

  17. [17]

    On Explainability of Graph Neural Networks via Subgraph Explorations , url =

    Hao Yuan and Haiyang Yu and Jie Wang and Kang Li and Shuiwang Ji , bibsource =. On Explainability of Graph Neural Networks via Subgraph Explorations , url =. Proceedings of the 38th International Conference on Machine Learning,

    work page

  18. [18]

    arXiv , author =:2107.11889 , primaryclass =

    work page arXiv

  19. [19]

    The Eleventh International Conference on Learning Representations,

    Wenqian Li and Yinchuan Li and Zhigang Li and Jianye Hao and Yan Pang , bibsource =. The Eleventh International Conference on Learning Representations,

    work page

  20. [20]

    Malliaros , journal =

    Alexandre Duval and Fragkiskos D. Malliaros , journal =. GraphSVX: Shapley Value Explanations for Graph Neural Networks , url =

    work page

  21. [21]

    GNNShap: Scalable and Accurate

    Selahattin Akkas and Ariful Azad , bibsource =. GNNShap: Scalable and Accurate. Proceedings of the. doi:10.1145/3589334.3645599 , editor =

    work page doi:10.1145/3589334.3645599

  22. [22]

    A Review on Graph Neural Network Methods in Financial Applications , url =

    Jianian Wang and Sheng Zhang and Yanghua Xiao and Rui Song , journal =. A Review on Graph Neural Network Methods in Financial Applications , url =

    work page

  23. [23]

    A Survey on Graph Neural Networks in Intelligent Transportation Systems , url =

    Hourun Li and Yusheng Zhao and Zhengyang Mao and Yifang Qin and Zhiping Xiao and Jiaqi Feng and Yiyang Gu and Wei Ju and Xiao Luo and Ming Zhang , journal =. A Survey on Graph Neural Networks in Intelligent Transportation Systems , url =

    work page

  24. [24]

    Kipf and Max Welling , bibsource =

    Thomas N. Kipf and Max Welling , bibsource =. Semi-Supervised Classification with Graph Convolutional Networks , url =. 5th International Conference on Learning Representations,

    work page

  25. [25]

    arXiv , author =:2210.17159 , primaryclass =

    work page arXiv

  26. [26]

    IEEE Transactions on Artificial Intelligence , number =

    Ragno, Alessio and La Rosa, Biagio and Capobianco, Roberto , doi =. IEEE Transactions on Artificial Intelligence , number =

    work page

  27. [27]

    Random Walk Graph Neural Networks , url =

    Giannis Nikolentzos and Michalis Vazirgiannis , bibsource =. Random Walk Graph Neural Networks , url =. Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual , editor =

    work page 2020

  28. [28]

    Concept Bottleneck Models , url =

    Pang Wei Koh and Thao Nguyen and Yew Siang Tang and Stephen Mussmann and Emma Pierson and Been Kim and Percy Liang , bibsource =. Concept Bottleneck Models , url =. Proceedings of the 37th International Conference on Machine Learning,

    work page

  29. [29]

    Learning Important Features Through Propagating Activation Differences , url =

    Avanti Shrikumar and Peyton Greenside and Anshul Kundaje , bibsource =. Learning Important Features Through Propagating Activation Differences , url =. Proceedings of the 34th International Conference on Machine Learning,

    work page

  30. [30]

    GraphFramEx: Towards Systematic Evaluation of Explainability Methods for Graph Neural Networks , url =

    Kenza Amara and Rex Ying and Zitao Zhang and Zhihao Han and Yinan Shan and Ulrik Brandes and Sebastian Schemm and Ce Zhang , journal =. GraphFramEx: Towards Systematic Evaluation of Explainability Methods for Graph Neural Networks , url =

    work page

  31. [31]

    Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks , url =

    Xu Zheng and Farhad Shirani and Tianchun Wang and Wei Cheng and Zhuomin Chen and Haifeng Chen and Hua Wei and Dongsheng Luo , bibsource =. Towards Robust Fidelity for Evaluating Explainability of Graph Neural Networks , url =. The Twelfth International Conference on Learning Representations,

    work page

  32. [32]

    Evaluating Explainability for Graph Neural Networks , url =

    Chirag Agarwal and Owen Queen and Himabindu Lakkaraju and Marinka Zitnik , journal =. Evaluating Explainability for Graph Neural Networks , url =

    work page

  33. [33]

    A Survey on Explainability of Graph Neural Networks , url =

    Jaykumar Kakkad and Jaspal Jannu and Kartik Sharma and Charu Aggarwal and Sourav Medya , journal =. A Survey on Explainability of Graph Neural Networks , url =

    work page

  34. [34]

    Improving Subgraph Recognition with Variational Graph Information Bottleneck , url =

    Junchi Yu and Jie Cao and Ran He , journal =. Improving Subgraph Recognition with Variational Graph Information Bottleneck , url =

    work page

  35. [35]

    GREASE: Generate Factual and Counterfactual Explanations for GNN-based Recommendations , url =

    Ziheng Chen and Fabrizio Silvestri and Jia Wang and Yongfeng Zhang and Zhenhua Huang and Hongshik Ahn and Gabriele Tolomei , journal =. GREASE: Generate Factual and Counterfactual Explanations for GNN-based Recommendations , url =

    work page

  36. [36]

    Robust Counterfactual Explanations on Graph Neural Networks , url =

    Mohit Bajaj and Lingyang Chu and Zi Yu Xue and Jian Pei and Lanjun Wang and Peter Cho. Robust Counterfactual Explanations on Graph Neural Networks , url =. Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual , editor =

    work page 2021

  37. [37]

    Interpretable and Generalizable Graph Learning via Stochastic Attention Mechanism , url =

    Siqi Miao and Mia Liu and Pan Li , bibsource =. Interpretable and Generalizable Graph Learning via Stochastic Attention Mechanism , url =. International Conference on Machine Learning,

    work page

  38. [38]

    Graph Attention Networks , url =

    Petar Velickovic and Guillem Cucurull and Arantxa Casanova and Adriana Romero and Pietro Li. Graph Attention Networks , url =. 6th International Conference on Learning Representations,

    work page

  39. [39]

    Sun and Krishna C

    Chris Lin and Gerald J. Sun and Krishna C. Bulusu and Jonathan R. Dry and Marylens Hernandez , journal =. Graph Neural Networks Including Sparse Interpretability , url =

    work page

  40. [40]

    arXiv preprint arXiv:2009.07896 , year=

    Captum: A unified and generic model interpretability library for PyTorch , year =. arXiv , author =:2009.07896 , primaryclass =

    work page arXiv 2009

  41. [41]

    Interpreting Graph Neural Networks for

    Michael Sejr Schlichtkrull and Nicola De Cao and Ivan Titov , bibsource =. Interpreting Graph Neural Networks for. 9th International Conference on Learning Representations,

    work page

  42. [42]

    A Survey of the State of Explainable

    Danilevsky, Marina and Qian, Kun and Aharonov, Ranit and Katsis, Yannis and Kawas, Ban and Sen, Prithviraj , booktitle =. A Survey of the State of Explainable

    work page

  43. [43]

    Explaining the Explainers in Graph Neural Networks: a Comparative Study , volume=

    Longa, Antonio and Azzolin, Steve and Santin, Gabriele and Cencetti, Giulia and Lio, Pietro and Lepri, Bruno and Passerini, Andrea , year=. Explaining the Explainers in Graph Neural Networks: a Comparative Study , volume=. ACM Computing Surveys , publisher=. doi:10.1145/3696444 , number=

    work page doi:10.1145/3696444

  44. [44]

    T., Singh, S., & Guestrin, C

    Nauta, Meike and Trienes, Jan and Pathak, Shreyasi and Nguyen, Elisa and Peters, Michelle and Schmitt, Yasmin and Schlötterer, Jörg and van Keulen, Maurice and Seifert, Christin , year=. From Anecdotal Evidence to Quantitative Evaluation Methods: A Systematic Review on Evaluating Explainable AI , volume=. ACM Computing Surveys , publisher=. doi:10.1145/35...

    work page doi:10.1145/3583558

  45. [45]

    2022 , eprint=

    How Attentive are Graph Attention Networks? , author=. 2022 , eprint=

    work page 2022

  46. [46]

    XGNN: Towards Model-Level Explanations of Graph Neural Networks , url=

    Yuan, Hao and Tang, Jiliang and Hu, Xia and Ji, Shuiwang , year=. XGNN: Towards Model-Level Explanations of Graph Neural Networks , url=. doi:10.1145/3394486.3403085 , booktitle=

    work page doi:10.1145/3394486.3403085

  47. [47]

    2025 , eprint=

    GSAT: Graph Structure Attention Networks , author=. 2025 , eprint=

    work page 2025

  48. [48]

    Learning and Evaluating Graph Neural Network Explanations based on Counterfactual and Factual Reasoning , url=

    Tan, Juntao and Geng, Shijie and Fu, Zuohui and Ge, Yingqiang and Xu, Shuyuan and Li, Yunqi and Zhang, Yongfeng , year=. Learning and Evaluating Graph Neural Network Explanations based on Counterfactual and Factual Reasoning , url=. doi:10.1145/3485447.3511948 , booktitle=

    work page doi:10.1145/3485447.3511948

  49. [49]

    2021 , eprint=

    Generative Causal Explanations for Graph Neural Networks , author=. 2021 , eprint=

    work page 2021

  50. [50]

    2022 , eprint=

    CF-GNNExplainer: Counterfactual Explanations for Graph Neural Networks , author=. 2022 , eprint=

    work page 2022

  51. [51]

    2023 , eprint=

    D4Explainer: In-Distribution GNN Explanations via Discrete Denoising Diffusion , author=. 2023 , eprint=

    work page 2023

  52. [52]

    2022 , eprint=

    GStarX: Explaining Graph Neural Networks with Structure-Aware Cooperative Games , author=. 2022 , eprint=

    work page 2022

  53. [53]

    2023 , eprint=

    DAG Matters! GFlowNets Enhanced Explainer For Graph Neural Networks , author=. 2023 , eprint=

    work page 2023

  54. [54]

    2021 , eprint=

    On Explainability of Graph Neural Networks via Subgraph Explorations , author=. 2021 , eprint=

    work page 2021

  55. [55]

    2022 , eprint=

    Zorro: Valid, Sparse, and Stable Explanations in Graph Neural Networks , author=. 2022 , eprint=

    work page 2022

  56. [56]

    Towards Multi-Grained Explainability for Graph Neural Networks , url =

    Wang, Xiang and Wu, Yingxin and Zhang, An and He, Xiangnan and Chua, Tat-Seng , booktitle =. Towards Multi-Grained Explainability for Graph Neural Networks , url =

    work page

  57. [57]

    Reinforced Causal Explainer for Graph Neural Networks , volume=

    Wang, Xiang and Wu, Yingxin and Zhang, An and Feng, Fuli and He, Xiangnan and Chua, Tat-Seng , year=. Reinforced Causal Explainer for Graph Neural Networks , volume=. IEEE Transactions on Pattern Analysis and Machine Intelligence , publisher=. doi:10.1109/tpami.2022.3170302 , number=

    work page doi:10.1109/tpami.2022.3170302 2022

  58. [58]

    2024 , eprint=

    GNNInterpreter: A Probabilistic Generative Model-Level Explanation for Graph Neural Networks , author=. 2024 , eprint=

    work page 2024

  59. [59]

    2025 , eprint=

    Robustness questions the interpretability of graph neural networks: what to do? , author=. 2025 , eprint=

    work page 2025

  60. [60]

    2020 , eprint=

    Graph Information Bottleneck , author=. 2020 , eprint=

    work page 2020

  61. [61]

    2024 , eprint=

    Learning to Extend Molecular Scaffolds with Structural Motifs , author=. 2024 , eprint=

    work page 2024

  62. [62]

    Vishwanathan and Nicol N

    S.V.N. Vishwanathan and Nicol N. Schraudolph and Risi Kondor and Karsten M. Borgwardt , title =. Journal of Machine Learning Research , year =

    work page

  63. [63]

    and Kolouri, Soheil and Rostami, Mohammad and Martin, Charles E

    Pope, Phillip E. and Kolouri, Soheil and Rostami, Mohammad and Martin, Charles E. and Hoffmann, Heiko , booktitle=. Explainability Methods for Graph Convolutional Neural Networks , year=

    work page

  64. [64]

    2022 , eprint=

    Explainability in Graph Neural Networks: A Taxonomic Survey , author=. 2022 , eprint=

    work page 2022

  65. [65]

    2017 , eprint=

    Axiomatic Attribution for Deep Networks , author=. 2017 , eprint=

    work page 2017

  66. [66]

    An Efficient Explanation of Individual Classifications using Game Theory , author=. J. Mach. Learn. Res. , year=

    work page

  67. [67]

    2022 , eprint=

    Towards Prototype-Based Self-Explainable Graph Neural Network , author=. 2022 , eprint=

    work page 2022