pith. sign in

arxiv: 2210.15304 · v3 · submitted 2022-10-27 · 💻 cs.LG · cs.AI

Explaining the Explainers in Graph Neural Networks: a Comparative Study

Antonio Longa , Steve Azzolin , Gabriele Santin , Giulia Cencetti , Pietro Li\`o , Bruno Lepri , Andrea Passerini This is my paper

Pith reviewed 2026-05-24 10:58 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords graph neural networksexplainabilityinterpretabilitybenchmarkingcomparative studynode classificationgraph classification
0
0 comments X

The pith

Systematic tests of ten explainers on eight GNN architectures and six datasets isolate the design choices that produce reliable interpretations.

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

The paper performs a controlled comparison of ten different methods for explaining the outputs of Graph Neural Networks. Eight representative GNN models are trained on six graph and node classification datasets, after which each explainer is evaluated using multiple standard metrics. The experiments move past simple rankings to identify which internal components of an explainer determine whether its outputs align with the model's actual decision process. The resulting guidelines help users select explainers appropriate to a given task and avoid common sources of misleading interpretations.

Core claim

The authors establish through direct experiment that explainer success depends primarily on specific implementation details such as the treatment of graph structure and the selection of reference baselines rather than on the explainer's broad category, and they extract practical rules for avoiding pitfalls including metric instability and mismatch between explainer assumptions and dataset properties.

What carries the argument

The comparative experimental protocol that holds architectures and datasets fixed while varying explainers to isolate which components drive faithful explanations.

If this is right

  • Explainers incorporating explicit graph-structure handling outperform generic adaptations in most tested node and graph classification settings.
  • Recommendations allow selection of an explainer according to whether the task is node classification or whole-graph classification.
  • Avoiding single-metric reliance and checking stability across similar inputs reduces the chance of accepting spurious explanations.
  • The study leaves open the question of how these component-level insights scale to larger graphs and additional architectures.

Where Pith is reading between the lines

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

  • The component-focused findings suggest that GNN training procedures could be modified to optimize directly for the identified successful explainer traits.
  • Extending the same protocol to attention-based or message-passing variants not included in the original eight architectures would test whether the same components remain decisive.
  • Domain scientists applying GNNs to chemistry or biology graphs may require additional validation steps before adopting the general selection rules.

Load-bearing premise

The ten chosen explainers, eight architectures, and six datasets are representative enough to support general recommendations about GNN explainability.

What would settle it

A new explainer or architecture outside the tested collection that produces explanations rated highly by the paper's metrics yet fails to align with model behavior on an additional dataset would falsify the generality of the derived recommendations.

Figures

Figures reproduced from arXiv: 2210.15304 by Andrea Passerini, Antonio Longa, Bruno Lepri, Gabriele Santin, Giulia Cencetti, Pietro Li\`o, Steve Azzolin.

Figure 1
Figure 1. Figure 1: An overview of the adopted GNN architectures structured in a taxonomy as defined by Zhou et al. [135]. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Explanation masks (node- or edge-based) computed by the different explainers on the predictions of Gcn. Each row visualizes the mask computed for a given random graph from each dataset. 7.2.2 RQ2: How do explainers affect the explanations? The answers to this question are summarized in [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Explanation masks computed by SubX on the predictions of the different models. Each row visualizes the mask computed for a given random graph from each dataset. 7.2.3 RQ3: How do different types of problems affect the explanations? To address the third question we analyze the datasets separately. We remark that each dataset has been chosen to represent different types of challenges, which will be discussed… view at source ↗
Figure 4
Figure 4. Figure 4: Left: fidelity and plausibility achieved by all the model-explainer pairs when applied to Grid. In each pair, the name refers to the model while the color identifies the explainer. Right: stability of the explanations for the three top-performing model-explanation pairs. The colors identify important edges (dark red), and the edge thickness the variability of the importance in the dataset. For these two to… view at source ↗
Figure 5
Figure 5. Figure 5: Examples of explanations provided for each model and its highest plausibility explainer, when applied to a random sample from Grid. The plausibility and fidelity values are those of the entire dataset, as reported in [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Left: Fidelity and plausibility achieved by all the model-explainer pairs when applied to Grid-House, for class 0 (left) and class 1 (right). In each pair the name refers to the model, while the color identifies the explainer. Right: Stability of the explanations for the three top-performing model-explanation pairs, for each of the two classes. The colors identify important edges (dark red), and the edge t… view at source ↗
Figure 7
Figure 7. Figure 7: Examples of explanations provided for each model and its highest plausibility explainer, when applied to a random sample from Grid-House. Each row shows the results for one of the two classes. The plausibility and fidelity values are those of the entire dataset, as reported in [PITH_FULL_IMAGE:figures/full_fig_p020_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Fidelity and plausibility achieved by all the model-explainer pairs when applied to Stars, for class 0 (left) and class 1 (right). In each pair the name refers to the model, while the color identifies the explainer [PITH_FULL_IMAGE:figures/full_fig_p021_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Stability of the explanations for the three top-performing model-explanation pairs, for each of the three classes. The colors identify important edges (dark red), and the edge thickness the variability of the importance in the dataset. we mean the GNN focusing on learning the discriminant features of class 0 (1) while predicting the remaining class 1 (0) by the absence of such discriminant feature, without… view at source ↗
Figure 10
Figure 10. Figure 10: Examples of explanations provided for each model and its highest plausibility explainer, when applied [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Left: fidelity and plausibility achieved by all the model-explainer pairs when applied to House-Color, where the two classes (0 and 1) are represented in the left and right panel. In each pair the name refers to the model, while the color identifies the explainer. The plot is limited to metrics larger than 0.5 to simplify the visualization. Right: stability of the explanations for the three top-performing… view at source ↗
Figure 12
Figure 12. Figure 12: Examples of explanations provided for each model and its highest plausibility explainer, when applied to a random sample from Grid-House. Each row shows the results for one of the two classes. The plausibility and fidelity values are those of the entire dataset, as reported in the right of [PITH_FULL_IMAGE:figures/full_fig_p024_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Explanation masks (node- or edge-based) computed by the different explainers on the predictions of GraphSage on Shapes and Infection. Each row visualizes the mask computed for a given random graph from each dataset. For each dataset, only the explainers which passed the filtering procedure are shown. 7.3.2 RQ2: How do explainers affect the explanations? [PITH_FULL_IMAGE:figures/full_fig_p025_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Explanation masks (node- or edge-based) computed on the predictions of [PITH_FULL_IMAGE:figures/full_fig_p026_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Fidelity and plausibility achieved by all the model-explainer pairs when applied to Shapes. In each pair the name refers to the model, while the color identifies the explainer. Class 0 is omitted from the visualization since it is less relevant for the discussion of the explainers. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Examples of explanations provided for each model by its highest plausibility explainer, when applied to a random node from Shapes. Each row shows the results for one of the three house-structure classes. The plausibility and fidelity values are those of the entire dataset, as reported in [PITH_FULL_IMAGE:figures/full_fig_p027_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Fidelity and plausibility achieved by all the model-explainer pairs when applied to Infection. In each pair the name refers to the model, while the color identifies the explainer. Example of explanations are shown in [PITH_FULL_IMAGE:figures/full_fig_p027_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Examples of explanations provided for each model by its highest plausibility explainer, when applied to a [PITH_FULL_IMAGE:figures/full_fig_p028_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Examples of Gcn explanations for all explainers on the Grid dataset. 43 [PITH_FULL_IMAGE:figures/full_fig_p043_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Examples of Gcn explanations for all explainers on the Grid-House dataset. 44 [PITH_FULL_IMAGE:figures/full_fig_p044_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Examples of Set2Set explanations for all explainers on the Stars dataset. 45 [PITH_FULL_IMAGE:figures/full_fig_p045_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: Examples of MinCutPool explanations for all explainers on the House-Color dataset. 46 [PITH_FULL_IMAGE:figures/full_fig_p046_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: Examples of GraphSage explanations for all explainers on the Shapes dataset. 47 [PITH_FULL_IMAGE:figures/full_fig_p047_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: Examples of GraphSage explanations for all explainers on the Infection dataset. 48 [PITH_FULL_IMAGE:figures/full_fig_p048_24.png] view at source ↗
read the original abstract

Following a fast initial breakthrough in graph based learning, Graph Neural Networks (GNNs) have reached a widespread application in many science and engineering fields, prompting the need for methods to understand their decision process. GNN explainers have started to emerge in recent years, with a multitude of methods both novel or adapted from other domains. To sort out this plethora of alternative approaches, several studies have benchmarked the performance of different explainers in terms of various explainability metrics. However, these earlier works make no attempts at providing insights into why different GNN architectures are more or less explainable, or which explainer should be preferred in a given setting. In this survey, we fill these gaps by devising a systematic experimental study, which tests ten explainers on eight representative architectures trained on six carefully designed graph and node classification datasets. With our results we provide key insights on the choice and applicability of GNN explainers, we isolate key components that make them usable and successful and provide recommendations on how to avoid common interpretation pitfalls. We conclude by highlighting open questions and directions of possible future research.

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 presents a systematic experimental benchmark of ten GNN explainers applied to eight GNN architectures across six graph and node classification datasets. It claims to isolate key components that determine explainer success, supply actionable recommendations for explainer choice, and highlight open questions in GNN explainability.

Significance. If the chosen experimental grid is representative, the work supplies a useful empirical map that goes beyond prior benchmarks by linking architectural properties to explainability outcomes and offering concrete guidance for practitioners. The systematic design and emphasis on isolating usable components are strengths.

major comments (2)
  1. [Section 3] Experimental setup (Section 3): the claim that the six datasets and eight architectures are 'representative' and 'carefully designed' to support general recommendations requires an explicit mapping showing coverage of homophily/heterophily, graph size, and feature-versus-structure signal regimes; without this mapping the isolated key components and resulting recommendations remain tied to the tested slice.
  2. [Results section (Tables 2-5)] Results section (Tables 2-5 and associated figures): performance differences used to derive recommendations are presented without reported standard deviations across random seeds or statistical significance tests, making it impossible to judge whether observed differences in explainer fidelity or stability are robust.
minor comments (2)
  1. [Abstract] Abstract: the summary of contributions mentions 'key insights' and 'recommendations' but does not preview any concrete quantitative outcomes or named pitfalls.
  2. [Section 2] Notation: the definitions of the explainability metrics (fidelity, stability, etc.) appear only after the experimental setup; moving a compact table of metric definitions to Section 2 would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments, which help strengthen the clarity of our experimental design and the robustness of our results. We address each major comment point by point below, indicating the revisions we will incorporate.

read point-by-point responses

  1. Referee: [Section 3] Experimental setup (Section 3): the claim that the six datasets and eight architectures are 'representative' and 'carefully designed' to support general recommendations requires an explicit mapping showing coverage of homophily/heterophily, graph size, and feature-versus-structure signal regimes; without this mapping the isolated key components and resulting recommendations remain tied to the tested slice.

    Authors: We agree that an explicit mapping would better substantiate our claims of representativeness. Although the datasets and architectures were selected to span a range of homophily levels, sizes, and signal regimes (as described in the original Section 3), we did not include a consolidated table. In the revision we will add a new table (or subsection) in Section 3 that explicitly maps each of the six datasets and eight architectures to homophily/heterophily, graph size categories, and feature-versus-structure dominance, thereby clarifying the coverage and supporting the general recommendations. revision: yes

  2. Referee: [Results section (Tables 2-5)] Results section (Tables 2-5 and associated figures): performance differences used to derive recommendations are presented without reported standard deviations across random seeds or statistical significance tests, making it impossible to judge whether observed differences in explainer fidelity or stability are robust.

    Authors: We concur that standard deviations and statistical tests are necessary to establish robustness. The current tables report mean performance only. In the revised manuscript we will rerun the experiments with multiple random seeds, add standard deviations to Tables 2-5 (and associated figures), and include appropriate statistical significance tests (e.g., paired t-tests or Wilcoxon signed-rank tests with p-values) for the key explainer comparisons used to derive recommendations. revision: yes

Circularity Check

0 steps flagged

Empirical benchmark study with no derivation chain

full rationale

The paper performs a systematic experimental comparison of ten GNN explainers across eight architectures and six datasets, reporting performance metrics and deriving recommendations directly from those results. No mathematical derivations, parameter fittings, predictions, or uniqueness theorems are present that could reduce to inputs by construction. Self-citations (if any) are used only for background on prior benchmarks and do not bear the load of the central claims. The work is self-contained as an empirical study whose findings are externally falsifiable via replication on the same or additional datasets.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The study rests on standard machine-learning benchmarking assumptions rather than new theoretical constructs.

axioms (1)
  • domain assumption The selected datasets and architectures capture the relevant variation in GNN explainability behavior.
    Generalization of findings depends on this representativeness claim stated in the abstract.

pith-pipeline@v0.9.0 · 5741 in / 1019 out tokens · 32662 ms · 2026-05-24T10:58:22.874129+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.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. xAI-Drop: Don't Use What You Cannot Explain

    cs.LG 2024-07 unverdicted novelty 5.0

    xAI-Drop introduces an explainability-based topological dropping regularizer for GNNs that outperforms state-of-the-art dropping methods in accuracy and explanation quality on real-world datasets.

Reference graph

Works this paper leans on

135 extracted references · 135 canonical work pages · cited by 1 Pith paper · 12 internal anchors

  1. [1]

    Evaluating explainability for graph neural networks

    Chirag Agarwal, Owen Queen, Himabindu Lakkaraju, and Marinka Zitnik. Evaluating explainability for graph neural networks. arXiv preprint arXiv:2208.09339, 2022

    work page arXiv 2022

  2. [2]

    Probing gnn explainers: A rigorous theoretical and empirical analysis of gnn explanation methods

    Chirag Agarwal, Marinka Zitnik, and Himabindu Lakkaraju. Probing gnn explainers: A rigorous theoretical and empirical analysis of gnn explanation methods. In International Conference on Artificial Intelligence and Statistics, pages 8969–8996. PMLR, 2022

    work page 2022

  3. [3]

    A two-stage gan for high-resolution retinal image generation and segmentation

    Paolo Andreini, Giorgio Ciano, Simone Bonechi, Caterina Graziani, Veronica Lachi, Alessandro Mecocci, Andrea Sodi, Franco Scarselli, and Monica Bianchini. A two-stage gan for high-resolution retinal image generation and segmentation. Electronics, 11(1):60, 2021

    work page 2021

  4. [4]

    Diffusion-convolutional neural networks

    James Atwood and Don Towsley. Diffusion-convolutional neural networks. In D. Lee, M. Sugiyama, U. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems, vol- ume 29. Curran Associates, Inc., 2016

    work page 2016

  5. [5]

    Global explainability of gnns via logic combination of learned concepts

    Steve Azzolin, Antonio Longa, Pietro Barbiero, Pietro Lio, and Andrea Passerini. Global explainability of gnns via logic combination of learned concepts. In The Eleventh International Conference on Learning Representations, 2022

    work page 2022

  6. [6]

    How to explain individual classification decisions.The Journal of Machine Learning Research, 11:1803– 1831, 2010

    David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and Klaus-Robert M¨ uller. How to explain individual classification decisions.The Journal of Machine Learning Research, 11:1803– 1831, 2010

    work page 2010

  7. [7]

    Neural Machine Translation by Jointly Learning to Align and Translate

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  8. [8]

    Explainability Techniques for Graph Convolutional Networks

    Federico Baldassarre and Hossein Azizpour. Explainability techniques for graph convolutional networks.arXiv preprint arXiv:1905.13686, 2019

    work page internal anchor Pith review Pith/arXiv arXiv 1905

  9. [9]

    Emergence of scaling in random networks.science, 286(5439):509– 512, 1999

    Albert-L´ aszl´ o Barab´ asi and R´ eka Albert. Emergence of scaling in random networks.science, 286(5439):509– 512, 1999

    work page 1999

  10. [10]

    Equivariant mesh attention networks

    Sourya Basu, Jose Gallego-Posada, Francesco Vigan` o, James Rowbottom, and Taco Cohen. Equivariant mesh attention networks. arXiv preprint arXiv:2205.10662, 2022

    work page arXiv 2022

  11. [11]

    Spectral clustering with graph neural networks for graph pooling

    Filippo Maria Bianchi, Daniele Grattarola, and Cesare Alippi. Spectral clustering with graph neural networks for graph pooling. In International Conference on Machine Learning, pages 874–883. PMLR, 2020

    work page 2020

  12. [12]

    Protein function prediction via graph kernels

    Karsten M Borgwardt, Cheng Soon Ong, Stefan Sch¨ onauer, SVN Vishwanathan, Alex J Smola, and Hans- Peter Kriegel. Protein function prediction via graph kernels. Bioinformatics, 21(suppl 1):i47–i56, 2005

    work page 2005

  13. [13]

    Towards Sparse Hierarchical Graph Classifiers

    C˘ at˘ alina Cangea, Petar Veliˇ ckovi´ c, Nikola Jovanovi´ c, Thomas Kipf, and Pietro Li` o. Towards sparse hierar- chical graph classifiers. arXiv preprint arXiv:1811.01287, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  14. [14]

    Khalil, Andrea Lodi, Christopher Morris, and Petar Veliˇ ckovi´ c

    Quentin Cappart, Didier Ch´ etelat, Elias B. Khalil, Andrea Lodi, Christopher Morris, and Petar Veliˇ ckovi´ c. Combinatorial optimization and reasoning with graph neural networks. In Zhi-Hua Zhou, editor, Proceedings of the Thirtieth International Joint Conference on Artificial Intelligence, IJCAI-21, pages 4348–4355. Inter- national Joint Conferences on...

    work page 2021

  15. [15]

    How faithful are self- explainable gnns? In The Second Learning on Graphs Conference, 2023

    Marc Christiansen, Lea Villadsen, Zhiqiang Zhong, Stefano Teso, and Davide Mottin. How faithful are self- explainable gnns? In The Second Learning on Graphs Conference, 2023

    work page 2023

  16. [16]

    High-Performance Neural Networks for Visual Object Classification

    Dan C Cire¸ san, Ueli Meier, Jonathan Masci, Luca M Gambardella, and J¨ urgen Schmidhuber. High- performance neural networks for visual object classification. arXiv preprint arXiv:1102.0183, 2011. 31

    work page internal anchor Pith review Pith/arXiv arXiv 2011

  17. [17]

    ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators

    Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. arXiv preprint arXiv:2003.10555, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2003

  18. [18]

    A comprehensive survey on trustworthy graph neural networks: Privacy, robustness, fairness, and explainability

    Enyan Dai, Tianxiang Zhao, Huaisheng Zhu, Jun Xu, Zhimeng Guo, Hui Liu, Jiliang Tang, and Suhang Wang. A comprehensive survey on trustworthy graph neural networks: Privacy, robustness, fairness, and explainability. ArXiv, abs/2204.08570, 2022

    work page arXiv 2022

  19. [19]

    Convolutional neural networks on graphs with fast localized spectral filtering

    Micha¨ el Defferrard, Xavier Bresson, and Pierre Vandergheynst. Convolutional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems, 29, 2016

    work page 2016

  20. [20]

    Eraser: A benchmark to evaluate rationalized nlp models

    Jay DeYoung, Sarthak Jain, Nazneen Fatema Rajani, Eric Lehman, Caiming Xiong, Richard Socher, and Byron C Wallace. Eraser: A benchmark to evaluate rationalized nlp models. arXiv preprint arXiv:1911.03429, 2019

    work page arXiv 1911

  21. [21]

    Higher-order clustering and pooling for graph neural networks

    Alexandre Duval and Fragkiskos Malliaros. Higher-order clustering and pooling for graph neural networks. In Proceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 426–435, 2022

    work page 2022

  22. [22]

    A generalization of transformer networks to graphs

    Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs. arXiv preprint arXiv:2012.09699, 2020

    work page arXiv 2012

  23. [23]

    On random graphs i.Publicationes mathematicae, 6(1):290–297, 1959

    Paul Erd˝ os and Alfr´ ed R´ enyi. On random graphs i.Publicationes mathematicae, 6(1):290–297, 1959

    work page 1959

  24. [24]

    Moghaddam, and Roger Wattenhofer

    Lukas Faber, Amin K. Moghaddam, and Roger Wattenhofer. When comparing to ground truth is wrong: On evaluating gnn explanation methods. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 332–341, 2021

    work page 2021

  25. [25]

    Simulate time-integrated coarse-grained molecular dynamics with geometric machine learning

    Xiang Fu, Tian Xie, Nathan J Rebello, Bradley D Olsen, and Tommi Jaakkola. Simulate time-integrated coarse-grained molecular dynamics with geometric machine learning. arXiv preprint arXiv:2204.10348, 2022

    work page arXiv 2022

  26. [26]

    Hard masking for explaining Graph Neural Networks, 2021

    Thorben Funke, Megha Khosla, and Avishek Anand. Hard masking for explaining Graph Neural Networks, 2021

    work page 2021

  27. [27]

    Z orro: Valid, sparse, and stable explanations in graph neural networks

    Thorben Funke, Megha Khosla, Mandeep Rathee, and Avishek Anand. Z orro: Valid, sparse, and stable explanations in graph neural networks. IEEE Transactions on Knowledge and Data Engineering, 2022

    work page 2022

  28. [28]

    Higher-order interaction goes neural: a substructure assembling graph attention network for graph classification

    Jianliang Gao, Jun Gao, Xiaoting Ying, Mingming Lu, and Jianxin Wang. Higher-order interaction goes neural: a substructure assembling graph attention network for graph classification. IEEE Transactions on Knowledge and Data Engineering, 2021

    work page 2021

  29. [29]

    Diffusion improves graph learning, 2022

    Johannes Gasteiger, Stefan Weißenberger, and Stephan G¨ unnemann. Diffusion improves graph learning, 2022

    work page 2022

  30. [30]

    Neural bipartite matching, 2020

    Dobrik Georgiev and Pietro Li` o. Neural bipartite matching, 2020

    work page 2020

  31. [31]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In Proceedings of the 34th International Conference on Machine Learning - Volume 70, page 1263–1272, 2017

    work page 2017

  32. [32]

    Neural message passing for quantum chemistry

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. Neural message passing for quantum chemistry. In International conference on machine learning, pages 1263–1272. PMLR, 2017

    work page 2017

  33. [33]

    Towards training gnns using explanation directed message passing

    Valentina Giunchiglia, Chirag Varun Shukla, Guadalupe Gonzalez, and Chirag Agarwal. Towards training gnns using explanation directed message passing. In Learning on Graphs Conference, pages 28–1. PMLR, 2022

    work page 2022

  34. [34]

    node2vec: Scalable feature learning for networks

    Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining, pages 855–864, 2016

    work page 2016

  35. [35]

    A survey of methods for explaining black box models

    Riccardo Guidotti, Anna Monreale, Salvatore Ruggieri, Franco Turini, Fosca Giannotti, and Dino Pedreschi. A survey of methods for explaining black box models. ACM computing surveys (CSUR), 51(5):1–42, 2018

    work page 2018

  36. [36]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017. 32

    work page 2017

  37. [37]

    Wavelets on graphs via spectral graph theory

    David K Hammond, Pierre Vandergheynst, and R´ emi Gribonval. Wavelets on graphs via spectral graph theory. Applied and Computational Harmonic Analysis, 30(2):129–150, 2011

    work page 2011

  38. [38]

    Neural Networks, pages 389–416

    Trevor Hastie, Robert Tibshirani, and Jerome Friedman. Neural Networks, pages 389–416. Springer New York, New York, NY, 2009

    work page 2009

  39. [39]

    Convolution kernels on discrete structures

    David Haussler et al. Convolution kernels on discrete structures. Technical report, Technical report, Depart- ment of Computer Science, University of California . . . , 1999

    work page 1999

  40. [40]

    Long short-term memory

    Sepp Hochreiter and J¨ urgen Schmidhuber. Long short-term memory. Neural Comput., 9(8):1735–1780, nov 1997

    work page 1997

  41. [41]

    Open graph benchmark: Datasets for machine learning on graphs

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems, 33:22118–22133, 2020

    work page 2020

  42. [42]

    Graphlime: Local interpretable model explanations for graph neural networks

    Qiang Huang, Makoto Yamada, Yuan Tian, Dinesh Singh, and Yi Chang. Graphlime: Local interpretable model explanations for graph neural networks. IEEE Transactions on Knowledge and Data Engineering, 2022

    work page 2022

  43. [43]

    Learning equivariant energy based models with equivariant stein variational gradient descent

    Priyank Jaini, Lars Holdijk, and Max Welling. Learning equivariant energy based models with equivariant stein variational gradient descent. Advances in Neural Information Processing Systems, 34:16727–16737, 2021

    work page 2021

  44. [44]

    Information retrieval test collections

    K Sparck Jones and Cornelis Joost Van Rijsbergen. Information retrieval test collections. Journal of documentation, 1976

    work page 1976

  45. [45]

    A survey on explain- ability of graph neural networks

    Jaykumar Kakkad, Jaspal Jannu, Kartik Sharma, Charu Aggarwal, and Sourav Medya. A survey on explain- ability of graph neural networks. arXiv preprint arXiv:2306.01958, 2023

    work page arXiv 2023

  46. [46]

    Moldata, a molecular bench- mark for disease and target based machine learning

    Arash Keshavarzi Arshadi, Milad Salem, Arash Firouzbakht, and Jiann Yuan. Moldata, a molecular bench- mark for disease and target based machine learning. Journal of Cheminformatics, 14, 03 2022

    work page 2022

  47. [47]

    Privacy and transparency in graph machine learning: A unified perspective

    Megha Khosla. Privacy and transparency in graph machine learning: A unified perspective. arXiv preprint arXiv:2207.10896, 2022

    work page arXiv 2022

  48. [48]

    Pure transformers are powerful graph learners

    Jinwoo Kim, Tien Dat Nguyen, Seonwoo Min, Sungjun Cho, Moontae Lee, Honglak Lee, and Seunghoon Hong. Pure transformers are powerful graph learners. arXiv preprint arXiv:2207.02505, 2022

    work page arXiv 2022

  49. [49]

    Semi-Supervised Classification with Graph Convolutional Networks

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  50. [50]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84–90, 2017

    work page 2017

  51. [51]

    Contributions to the Theory of Games (AM-28), Volume II

    Harold William Kuhn and Albert William Tucker, editors. Contributions to the Theory of Games (AM-28), Volume II. Princeton University Press, Princeton, 1953

    work page 1953

  52. [52]

    Self-attention graph pooling

    Junhyun Lee, Inyeop Lee, and Jaewoo Kang. Self-attention graph pooling. In International conference on machine learning, pages 3734–3743. PMLR, 2019

    work page 2019

  53. [53]

    Improving semantic dependency parsing with higher-order information encoded by graph neural networks

    Bin Li, Yunlong Fan, Yikemaiti Sataer, Zhiqiang Gao, and Yaocheng Gui. Improving semantic dependency parsing with higher-order information encoded by graph neural networks. Applied Sciences, 12(8):4089, 2022

    work page 2022

  54. [54]

    Higher-order attribute-enhancing heterogeneous graph neural networks

    Jianxin Li, Hao Peng, Yuwei Cao, Yingtong Dou, Hekai Zhang, Philip Yu, and Lifang He. Higher-order attribute-enhancing heterogeneous graph neural networks. IEEE Transactions on Knowledge and Data Engineering, 2021

    work page 2021

  55. [55]

    Explainability in graph neural networks: An experimental survey

    Peibo Li, Yixing Yang, Maurice Pagnucco, and Yang Song. Explainability in graph neural networks: An experimental survey. arXiv preprint arXiv:2203.09258, 2022

    work page arXiv 2022

  56. [56]

    Rethinking graph neural networks for the graph coloring problem, 2022

    Wei Li, Ruxuan Li, Yuzhe Ma, Siu On Chan, David Pan, and Bei Yu. Rethinking graph neural networks for the graph coloring problem, 2022

    work page 2022

  57. [57]

    Dag matters! gflownets enhanced explainer for graph neural networks, 2023

    Wenqian Li, Yinchuan Li, Zhigang Li, Jianye Hao, and Yan Pang. Dag matters! gflownets enhanced explainer for graph neural networks, 2023. 33

    work page 2023

  58. [58]

    A survey of explainable graph neural networks: Taxonomy and evaluation metrics

    Yiqiao Li, Jianlong Zhou, Sunny Verma, and Fang Chen. A survey of explainable graph neural networks: Taxonomy and evaluation metrics. arXiv preprint arXiv:2207.12599, 2022

    work page arXiv 2022

  59. [59]

    Generative causal explanations for graph neural networks

    Wanyu Lin, Hao Lan, and Baochun Li. Generative causal explanations for graph neural networks. In International Conference on Machine Learning, pages 6666–6679. PMLR, 2021

    work page 2021

  60. [60]

    Orphicx: A causality-inspired latent variable model for interpreting graph neural networks

    Wanyu Lin, Hao Lan, Hao Wang, and Baochun Li. Orphicx: A causality-inspired latent variable model for interpreting graph neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13729–13738, 2022

    work page 2022

  61. [61]

    A survey of visual transformers

    Yang Liu, Yao Zhang, Yixin Wang, Feng Hou, Jin Yuan, Jiang Tian, Yang Zhang, Zhongchao Shi, Jianping Fan, and Zhiqiang He. A survey of visual transformers. arXiv preprint arXiv:2111.06091, 2021

    work page arXiv 2021

  62. [62]

    Graph neural networks for temporal graphs: State of the art, open challenges, and opportunities

    Antonio Longa, Veronica Lachi, Gabriele Santin, Monica Bianchini, Bruno Lepri, Pietro Lio, franco scarselli, and Andrea Passerini. Graph neural networks for temporal graphs: State of the art, open challenges, and opportunities. Transactions on Machine Learning Research, 2023

    work page 2023

  63. [63]

    Ter Hoeve, Gabriele Tolomei, Maarten De Rijke, and Fabrizio Silvestri

    Ana Lucic, Maartje A. Ter Hoeve, Gabriele Tolomei, Maarten De Rijke, and Fabrizio Silvestri. Cf- gnnexplainer: Counterfactual explanations for graph neural networks. In Gustau Camps-Valls, Francisco J. R. Ruiz, and Isabel Valera, editors, Proceedings of The 25th International Conference on Artificial Intelligence and Statistics, volume 151 of Proceedings ...

    work page 2022

  64. [64]

    Param- eterized explainer for graph neural network

    Dongsheng Luo, Wei Cheng, Dongkuan Xu, Wenchao Yu, Bo Zong, Haifeng Chen, and Xiang Zhang. Param- eterized explainer for graph neural network. Advances in neural information processing systems, 33:19620– 19631, 2020

    work page 2020

  65. [65]

    On data-aware global explainability of graph neural networks

    Ge Lv and Lei Chen. On data-aware global explainability of graph neural networks. Proc. VLDB Endow., 16(11):3447–3460, aug 2023

    work page 2023

  66. [66]

    On glocal explainability of graph neural networks

    Ge Lv, Lei Chen, and Caleb Chen Cao. On glocal explainability of graph neural networks. In Database Systems for Advanced Applications: 27th International Conference, DASFAA 2022, Virtual Event, April 11–14, 2022, Proceedings, Part I, page 648–664, Berlin, Heidelberg, 2022. Springer-Verlag

    work page 2022

  67. [67]

    Encoding concepts in graph neural networks

    Lucie Charlotte Magister, Pietro Barbiero, Dmitry Kazhdan, Federico Siciliano, Gabriele Ciravegna, Fabrizio Silvestri, Mateja Jamnik, and Pietro Lio. Encoding concepts in graph neural networks. arXiv preprint arXiv:2207.13586, 2022

    work page arXiv 2022

  68. [68]

    Gcexplainer: Human-in-the-loop concept-based explanations for graph neural networks, 2021

    Lucie Charlotte Magister, Dmitry Kazhdan, Vikash Singh, and Pietro Li` o. Gcexplainer: Human-in-the-loop concept-based explanations for graph neural networks, 2021

    work page 2021

  69. [69]

    Interpretable and generalizable graph learning via stochastic attention mechanism

    Siqi Miao, Mia Liu, and Pan Li. Interpretable and generalizable graph learning via stochastic attention mechanism. In International Conference on Machine Learning, pages 15524–15543. PMLR, 2022

    work page 2022

  70. [70]

    Discrete-time dynamic graph echo state networks

    Alessio Micheli and Domenico Tortorella. Discrete-time dynamic graph echo state networks. Neurocomputing, 496:85–95, 2022

    work page 2022

  71. [71]

    Weisfeiler and leman go neural: Higher-order graph neural networks

    Christopher Morris, Martin Ritzert, Matthias Fey, William L Hamilton, Jan Eric Lenssen, Gaurav Rattan, and Martin Grohe. Weisfeiler and leman go neural: Higher-order graph neural networks. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages 4602–4609, 2019

    work page 2019

  72. [72]

    Graphchef: Decision-tree recipes to explain graph neural networks

    Peter M¨ uller, Lukas Faber, Karolis Martinkus, and Roger Wattenhofer. Graphchef: Decision-tree recipes to explain graph neural networks. In The Twelfth International Conference on Learning Representations, 2024

    work page 2024

  73. [73]

    Globally interpretable graph learning via distribution matching

    Yi Nian, Yurui Chang, Wei Jin, and Lu Lin. Globally interpretable graph learning via distribution matching. In Proceedings of the ACM on Web Conference 2024, pages 992–1002, 2024

    work page 2024

  74. [74]

    Review of tools and algorithms for network motif discovery in biological networks

    Sabyasachi Patra and Anjali Mohapatra. Review of tools and algorithms for network motif discovery in biological networks. IET systems biology, 14(4):171–189, 2020

    work page 2020

  75. [75]

    Explainability methods for graph convolutional neural networks

    Phillip E Pope, Soheil Kolouri, Mohammad Rostami, Charles E Martin, and Heiko Hoffmann. Explainability methods for graph convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10772–10781, 2019. 34

    work page 2019

  76. [76]

    Marcelo Prates, Pedro H. C. Avelar, Henrique Lemos, Luis C. Lamb, and Moshe Y. Vardi. Learning to solve np-complete problems: A graph neural network for decision tsp. In Proceedings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Conference and Ninth AAAI Symposium on Educa...

    work page 2019

  77. [77]

    Frame averaging for invariant and equivariant network design

    Omri Puny, Matan Atzmon, Heli Ben-Hamu, Edward J Smith, Ishan Misra, Aditya Grover, and Yaron Lipman. Frame averaging for invariant and equivariant network design. arXiv preprint arXiv:2110.03336, 2021

    work page arXiv 2021

  78. [78]

    Exploring the limits of transfer learning with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J Liu, et al. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(140):1–67, 2020

    work page 2020

  79. [79]

    Prototype-based interpretable graph neural net- works

    Alessio Ragno, Biagio La Rosa, and Roberto Capobianco. Prototype-based interpretable graph neural net- works. IEEE Transactions on Artificial Intelligence, 5(4):1486–1495, 2022

    work page 2022

  80. [80]

    Hierarchical Text-Conditional Image Generation with CLIP Latents

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022

    work page internal anchor Pith review Pith/arXiv arXiv 2022

Showing first 80 references.