Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper proves that sparsity- and information-bottleneck-trained self-explainable GNNs extract Minimal Explanations, that these coincide with Prime Implicant explanations only for positive existential (motif) classifiers, and that they…

desk verdict Useful formal analysis of SE-GNN explanations, but Theorem 4.2's ME-PI equivalence needs a definitional fix before it is as strong as claimed. read the letter →

arxiv 2502.02719 v2 pith:KCR5CXHM submitted 2025-02-04 cs.LG

Steve Azzolin , Sagar Malhotra , Andrea Passerini , Stefano Teso This is my paper · ORCID
classification cs.LG MSC 68T0768T30
keywords self-explainablegraphneuralnetworksminimalexplanationsprimeimplicantfaithfulnessnetworkexplainabilityfirst-orderlogicclassifiersdual-channelGNNsinformationbottleneck
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

Self-Explainable Graph Neural Networks (SE-GNNs) are marketed as interpretable-by-design: they extract a subgraph of the input and classify from it. This paper asks what kind of explanation that subgraph actually is, and answers with a formal characterization plus a fix. It proves that, under idealized conditions (a ground-truth classifier, a hard extractor, perfect accuracy), the training losses used by popular sparsity- and information-bottleneck-based SE-GNNs are minimized exactly when the extracted subgraph is a Minimal Explanation, i.e., the smallest subgraph that preserves the prediction. It then shows that Minimal Explanations coincide with Prime Implicant explanations—minimal subgraphs whose every supergraph preserves the prediction—for classifiers expressible as purely existentially quantified positive first-order logic formulas, which covers motif-recognition tasks. For other tasks, Minimal Explanations can be identical for two genuinely different classifiers and can achieve zero faithfulness, so the paper proposes Dual-Channel GNNs that add a simple white-box rule channel and let the model decide whether to use topology, rules, or both.

What carries the argument

The load-bearing objects are two formal explanation notions defined on subgraphs. A Minimal Explanation (ME) for prediction $g(G)$ is a smallest subgraph $R \subseteq G$ with $g(R)=g(G)$. A Prime Implicant (PI) explanation is a subgraph $R$ such that every supergraph of $R$ inside $G$ keeps the prediction, and no strict subgraph has that property; PIs are minimally sufficient explanations. Theorem 3.2 connects these to training: with a ground-truth classifier $f$, a hard extractor $q$ (scores in $\{0,1\}$, or $\{r,1\}$ for IB losses), and perfect accuracy, the sparsity and information-bottleneck objectives of Table 1 are minimal iff $q(G)$ is an ME. Theorem 4.2 is the positive bridge: for purely existentially quantified positive FOL classifiers, every ME on a positive instance is also a PI, because any supergraph still contains the witnessing tuple and any smaller subgraph loses it. The negative results ride on the two example classifiers $\exists x \exists y. E(x,y)$ and $\forall x \exists y. E(x,y)$, which have identical MEs but different PIs; the faithfulness definitions Suf and Nec (Definition 5.1) yield Propositions 5.3 and 5.4, which show MEs maximize sufficiency only when they contain a PI and need non-zero necessity only if they intersect every PI. Finally, the proposed Dual-Channel GNN architecture replaces some of the work of the subgraph channel with a sparse linear rule channel $g_2(G)=\sigma(W \sum_{u} x_u)$ combined through a temperature-annealed logic network, letting the model choose topology, rule, or both.

What would settle it

Take a triangle under the classifier $\forall x \exists y. E(x,y)$: a single edge is a Minimal Explanation but a Prime Implicant explanation is an edge cover of all three vertices, so the ME-equals-PI claim fails outside the positive existential fragment—this is the paper's own boundary example. To test Theorem 3.2 concretely, enumerate all subgraphs of a small graph and verify that any perfectly accurate hard-extractor SE-GNN with minimal Table 1 loss always outputs a smallest label-preserving subgraph; one counterexample with a non-minimal subgraph would refute the if-and-only-if.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that the explanation type that sparsity- and IB-based SE-GNNs optimize is a Minimal Explanation, not a Prime Implicant or a faithful explanation, and that these three notions separate cleanly. Theorem 3.2 states the if-and-only-if: an idealized SE-GNN with a ground-truth classifier, hard saturated scores, and perfect accuracy minimizes the Table 1 risks exactly when $q(G)$ is a Minimal Explanation for every instance. Theorem 4.2 then shows that for any purely existentially quantified positive first-order logic classifier, every Minimal Explanation on a positive instance is also a Prime Implicant explanation, so the explanations inherit sufficiency and minimality for motif-based tasks. Theorem 3.4 and Theorem 5.2 show the limits: two distinct classifiers can share all Minimal Explanations on every co-labeled instance (e.g., edge-existence versus no-isolated-nodes), and a Minimal Explanation can have zero faithfulness because deleting it need not change the label when other evidence remains. The constructive response is Dual-Channel GNNs, which pair the subgraph channel with a sparse linear rule channel and adaptively weight the two; the experiments report that this simple pairing matches or improves accuracy while yielding more compact subgraph explanations and extractable rules.

Load-bearing premise

The theoretical equivalence assumes the model has a perfect, ground-truth classifier, an explanation extractor whose scores saturate to hard values, and perfect predictive accuracy; real trained models use soft scores and finite capacity, so the minimal-explanation guarantee describes an idealized optimum rather than a property of every trained network.

Editorial extensions

If this is right

  • On motif-based tasks expressible as purely existentially quantified positive first-order logic formulas, the subgraphs extracted by SE-GNNs are Prime Implicant explanations, so they are minimally sufficient and inherit the desirable properties of PIs.
  • In general, two distinct classifiers can produce identical Minimal Explanations on every instance where they agree, so local (and aggregated model-level) ME-based explanations do not always reveal which classifier is being explained.
  • A Minimal Explanation can have zero faithfulness: for an edge-existence classifier on a multi-edge graph, deleting the single explained edge leaves other edges, so necessity is zero and the overall faithfulness score is zero.
  • Prime Implicant and faithful explanations can be as large as the input (edge covers, or the whole graph) and are intractable to find, which argues against making them the training target directly.
  • Dual-Channel GNNs, which pair the topological SE-GNN with a sparse linear rule channel, match or improve accuracy on the tested datasets and can recover succinct rules such as 'number of red nodes at least two' while keeping subgraph explanations more compact.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the ME/PI gap persists in soft-score trained models, then faithfulness-based benchmarks will systematically penalize SE-GNNs on any task with redundant evidence; a quick test is to measure necessity of a one-edge explanation on a multi-edge graph and check whether it drops to zero.
  • Editorial inference: the DC-GNN design suggests a reusable recipe—offload global or feature-based conditions to a symbolic rule channel while leaving motifs to the subgraph explainer; molecular property tasks with both scaffold motifs and global features such as size or charge are a natural testbed.
  • Editorial inference: the positive existential boundary is likely not the only fragment with ME equals PI; mapping the exact fragment of first-order logic where the equality holds would turn the boundary result into a classification of tasks where subgraph explanations are safe.
  • Editorial inference: because the aggregation gate is itself interpretable, DC-GNNs could support interactive debugging—a user who sees the model rely on the rule channel knows the subgraph explanation is not where the decision lives.
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.

Referee Report

3 major / 4 minor

Summary. The paper formalizes the explanations produced by sparsity- and information-bottleneck-based self-explainable GNNs as Minimal Explanations (MEs) and compares them with Prime Implicant (PI) and faithful explanations. It proves that, under idealized assumptions (ground-truth classifier, hard extractor, perfect predictive accuracy), the considered SE-GNN losses are minimized exactly when the extracted subgraph is a ME; that MEs coincide with PIs for purely existential positive first-order-logic classifiers; that MEs can be uninformative or unfaithful in general; and that PIs can be prohibitively large. It then proposes Dual-Channel GNNs, which combine a subgraph-based SE-GNN channel with a sparse linear rule channel through an interpretable aggregation, and reports experiments on synthetic and real datasets showing competitive accuracy, adaptive channel selection, and more compact or faithful explanations. Code is publicly available.

Significance. If corrected, this is a useful and timely contribution: it gives one of the first formal bridges between SE-GNN explanation objectives and logic-based notions of explanation, and it provides a concrete architectural remedy (DC-GNNs) that is validated by ablations, OOD splits, and faithfulness measurements. The theoretical results are developed from first principles against external definitions rather than fitted quantities, and the empirical evaluation includes ground-truth rule checks in synthetic datasets and public code. The main technical claim, however, is currently over-stated for the manuscript's own general definitions, so the formal part needs a targeted revision before the contribution is fully sound.

major comments (3)
  1. [Definition 3.1; Definition 4.1; Theorem 4.2; Appendix A.2] The proof of Theorem 4.2 silently strengthens ME minimality: Definition 3.1 forbids only subgraphs R' with |R'| < |R|, while the proof ("there exists no |R'| ≤ |R|") forbids equal-size proper subgraphs. This is load-bearing. For |·| equal to the number of edges, take G to be a triangle plus an isolated node c and g = ∃x∃y.E(x,y); R = {(u,v)} ∪ {c} has size 1 and is a ME, because no 0-edge subgraph of G is positive. But R is not a PI, since R' = {(u,v)} is a proper subgraph of R with the same size that satisfies the PI sufficiency condition. The statement "a ME is induced by nodes in a tuple ā" is an extra assumption not contained in Definition 3.1. The theorem and the following paragraph therefore overclaim for general subgraphs; the result is recoverable for edge-induced (or node-induced) subgraphs with the corresponding cardinality, which is the actual SE-GNN setting in Section 2 and Table 1. Please add this restriction to the formal statements or redefine ME minimality consistently.
  2. [Proposition 4.3; Appendix A.3] The proof of Proposition 4.3 asserts that for a ME R, any proper R' ⊂ R satisfies g(R') ≠ g(G). This is the same equality-case gap: ME minimality only excludes strictly smaller subgraphs. Using the triangle-plus-isolated-node example from the previous comment, R = {(u,v)} ∪ {c} is a ME for ∃x∃y.E(x,y) but R' = {(u,v)} has the same edge count and the same label, so the claimed inclusion ∪_G ME(g(G)) ⊆ ∪_G PI(g(G)) is false for general subgraphs. The proposition needs the same edge-induced or node-induced restriction, or a genuinely different proof.
  3. [Definition 5.1; Proposition 5.3] Proposition 5.3 is stated for an arbitrary perturbation distribution pR, but its proof assumes that Suf(R)=1 is equivalent to every extension of R in G preserving the label. If pR is the degenerate distribution that never perturbs, every R has maximal Suf, so the "only if" direction fails. Please state explicitly that pR has full support over the relevant edge-deletion (or completion) set, or restrict the proposition to the perturbation family used in the faithfulness experiments.
minor comments (4)
  1. [Theorem 3.4 proof] In the proof of Theorem 3.4, the equalities ME(g(G)) = ME(g′(G)) = E and = V conflate the edge and node sets with sets of single-edge and single-node subgraphs; the intended statement is ME(g(G)) = {{e} : e ∈ E} and similarly for V.
  2. [Appendix A.3] The phrase "any extension of R (within R)" in the proof of Proposition 4.3 is confusing; the only extension of R inside R is R itself, so the sentence should say that condition (2) holds vacuously.
  3. [Table 3] The footnote explaining the '*' marker is hard to parse; please state explicitly that the marker denotes the channel chosen in nine out of ten seeds and that the exceptional seed is discussed in the text.
  4. [Section 2] The notation R ⊂ G versus R ⊆ G is introduced informally; a precise convention for whether subgraphs inherit all, some, or none of the node and edge features would help avoid ambiguity in Definition 3.1 and Definition 4.1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's formal results are derived from stated definitions and its empirical claims are benchmarked against ground-truth rules.

full rationale

The central formal results are not circular. Theorem 3.2 establishes an equivalence between the regularized training objectives of Table 1 and Definition 3.1; the proof reduces the loss to L(f(q(G)),Y)+λ1|q(G)|/|E| and observes that under perfect accuracy and hard extractors the minimizer is the smallest label-preserving subgraph, which is exactly the definition of a Minimal Explanation. This is a formal restatement rather than a fitted prediction, and it does not smuggle the conclusion into the inputs. Theorems 3.4, 4.4, 5.2 and Propositions 5.3-5.4 are proven directly from the definitions of ME, PI and faithfulness, with no fitted parameter renamed as a prediction. The empirical evaluation of DC-GNNs is self-contained: channel selection and extracted rules are read off trained models and compared with ground-truth generation rules, and the theoretical ME/PI claims are not fitted to the experimental data. Self-citations (e.g., Azzolin et al. 2025 for the faithfulness definition and readout mitigation) are not load-bearing because the cited definition is stated in full and the results follow from it. For completeness, the proof of Theorem 4.2 (Appendix A.2) contains a correctness gap: it silently upgrades the ME minimality condition from |R'| < |R| to |R'| ≤ |R| and assumes a ME is induced by a witnessing tuple, which is not in Definition 3.1. This is a proof error in the general-subgraph formulation, not a circularity, because the theorem's conclusion is not an input to the argument by construction.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central theoretical claim (Theorem 3.2) is self-contained and uses no fitted parameters. The main axioms are the idealization assumptions about SE-GNN training (perfect accuracy, hard extractor, ground-truth classifier) and the choice of faithfulness metric. No new entities are postulated.

assumptions (5)
  • domain assumption The SE-GNN has perfect predictive accuracy and its explanation extractor is hard (edge scores saturate to {0,1} or {r,1}).
    Stated in Section 3 and Appendix A.1 before Theorem 3.2; without this, the loss does not reduce to minimizing |q(G)|.
  • domain assumption The classifier f in the SE-GNN expresses the ground truth function.
    Used in the proof of Theorem 3.2 (Appendix A.1) to guarantee L(f(q(G)),Y) is minimized exactly.
  • domain assumption The graph classifiers used in examples (∃x∃y.E(x,y), ∀x∃y.E(x,y)) are expressible by GNNs.
    Invoked in Section 2 via Grohe (2021) Theorem IX.3 (C2 fragment).
  • domain assumption Faithfulness is measured as in Definition 5.1 with the exponential normalization from Azzolin et al. (2025).
    The zero-faithfulness result in Theorem 5.2 depends on this specific metric choice.
  • standard math Subgraph size |R| is a well-defined measure (nodes, edges, or features) and minimization is with respect to that measure.
    Definition 3.1 requires a total order on subgraph sizes; the results hold for any monotone size measure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective." pith.science (2026). https://pith.science/paper/KCR5CXHM

@misc{pith2026250202719,
  author       = {Pith},
  title        = {Pith review of: Beyond Topological Self-Explainable GNNs: A Formal Explainability Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KCR5CXHM}},
  note         = {Machine review of arXiv:2502.02719}
}
read the original abstract

Self-Explainable Graph Neural Networks (SE-GNNs) are popular explainable-by-design GNNs, but their explanations' properties and limitations are not well understood. Our first contribution fills this gap by formalizing the explanations extracted by some popular SE-GNNs, referred to as Minimal Explanations (MEs), and comparing them to established notions of explanations, namely Prime Implicant (PI) and faithful explanations. Our analysis reveals that MEs match PI explanations for a restricted but significant family of tasks. In general, however, they can be less informative than PI explanations and are surprisingly misaligned with widely accepted notions of faithfulness. Although faithful and PI explanations are informative, they are intractable to find and we show that they can be prohibitively large. Given these observations, a natural choice is to augment SE-GNNs with alternative modalities of explanations taking care of SE-GNNs' limitations. To this end, we propose Dual-Channel GNNs that integrate a white-box rule extractor and a standard SE-GNN, adaptively combining both channels. Our experiments show that even a simple instantiation of Dual-Channel GNNs can recover succinct rules and perform on par or better than widely used SE-GNNs.

Figures

Figures reproduced from arXiv: 2502.02719 by the authors.

Figure 1
Figure 1. Examples of a Minimal, PI, and faithful explana [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of our Dual-Channel architecture for a positive instance of TopoFeature, where positive instances contain a cycle and at least two red nodes. Numbers indicate edge relevance scores. SE-GNNs may fail to highlight precisely the elements relevant to the prediction, as the task involves non-topological patterns. DC-GNNs, instead, provide more focused topological explanations by offloading part of the predic… view at source ↗
Figure 3
Figure 3. Decision boundary of the linear classifier of [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Decision boundary of the linear classifier of [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]
Figure 5
Figure 5. Figure 5: Histograms of explanation relevance scores for GSAT on TopoFeature (validation set). The model fails to reliably separate between relevant and non-relevant edges, making it difficult to select a proper relevance threshold. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_5.png]
Figure 6
Figure 6. Figure 6: Histograms of explanation relevance scores for SMGNN on TopoFeature (validation set). The sparsification mechanism of SMGNN better separaMEs edges with higher importance than the rest of the graph. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_6.png]
Figure 7
Figure 7. Figure 7: Examples of explanations for SMGNN (seed 1) over TopoFeature. Relevant edges are those with puv ≥ 0.8 and are highlighted in red. Edges are annotated with their respective puv score. Samples of class 1 must have both a cycle and at least 2 red nodes. 27 [PITH_FULL_IMA…
Figure 8
Figure 8. Figure 8: Histograms of explanation relevance scores for DC-SMGNN on TopoFeature (validation set). Since the underlying SMGNN is now only looking for the topological motif (as the rule at least two red nodes is learned by the interpretable model), the Dual-Channel GNN is allowed…
Figure 9
Figure 9. Figure 9: Histograms of explanation relevance scores for DC-GSAT on TopoFeature (validation set). 29 [PITH_FULL_IMAGE:figures/full_fig_p029_9.png]
Figure 10
Figure 10. Figure 10: Examples of explanations for DC-SMGNN (seed 1) over TopoFeature. Relevant edges are those with puv ≥ 0.2 and are highlighted in red. The threshold is picked by looking at the histogram in [PITH_FULL_IMAGE:figures/full_fig_p030_10.png]
Figure 11
Figure 11. Figure 11: Histograms of explanation relevance scores for GSAT on RedBlueNodes (validation set). The model fails to reliably separate between relevant and non-relevant edges, making it difficult to select a proper relevance threshold. Specifically, most edges are assigned an imp…
Figure 12
Figure 12. Figure 12: Histograms of explanation relevance scores for SMGNN on RedBlueNodes ((validation set)). The model assigns very cluttered scores to almost all edges, failing to highlight a subset that is reliably more relevant than the others, making it difficult to select an appropr…
Figure 13
Figure 13. Figure 13: Examples of explanations for GSAT (seed 1) over RedBlueNodes. Relevant edges are those with puv ≥ 0.7 and are highlighted in red. Edges are not annotated with their respective puv score to avoid excessive clutter. Idx: 3 Class 1 (a) Idx: 10 Class 0 (b) [PITH_FULL_IMA…
Figure 14
Figure 14. Figure 14: Examples of explanations for SMGNN (seed 1) over RedBlueNodes. Relevant edges are those with puv ≥ 0.2 and are highlighted in red. Edges are not annotated with their respective puv score to avoid excessive clutter. 33 [PITH_FULL_IMAGE:figures/full_fig_p033_14.png]
Figure 15
Figure 15. Figure 15: Histograms of explanation relevance scores for GSAT on AIDS (validation set). 34 [PITH_FULL_IMAGE:figures/full_fig_p034_15.png]
Figure 16
Figure 16. Figure 16: Examples of explanations for GSAT (seed 8) over AIDS. Relevant edges are those with puv ≥ 0.8 and are highlighted in red. Edges are not annotated with their respective puv score to avoid excessive clutter. 35 [PITH_FULL_IMAGE:figures/full_fig_p035_16.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations

    cs.LG 2025-06 conditional novelty 6.0 of 10

    Abstraction-refinement over neuron merging computes provably sufficient and minimal explanations of neural network predictions substantially faster than verifying on the full network.

Reference graph

Works this paper leans on

84 extracted references · 54 canonical work pages · cited by 1 Pith paper

  1. [1]

    Evaluating explainability for graph neural networks

    Agarwal, C., Queen, O., Lakkaraju, H., and Zitnik, M. Evaluating explainability for graph neural networks. Scientific Data, 10 0 (1): 0 144, 2023

    2023

  2. [2]

    H., and Lakkaraju, H

    Agarwal, C., Tanneru, S. H., and Lakkaraju, H. Faithfulness vs. plausibility: On the (un)reliability of explanations from large language models, 2024

    2024

  3. [3]

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

    Amara, K., Ying, Z., Zhang, Z., Han, Z., Zhao, Y., Shan, Y., Brandes, U., Schemm, S., and Zhang, C. Graphframex: Towards systematic evaluation of explainability methods for graph neural networks. In Learning on Graphs Conference, pp.\ 44--1. PMLR, 2022

    2022

  4. [4]

    Global explainability of gnns via logic combination of learned concepts

    Azzolin, S., Longa, A., Barbiero, P., Lio, P., and Passerini, A. Global explainability of gnns via logic combination of learned concepts. In The Eleventh International Conference on Learning Representations, 2022

    2022

  5. [5]

    Reconsidering Faithfulness in Regular, Self-Explainable and Domain Invariant GNNs

    Azzolin, S., Longa, A., Teso, S., and Passerini, A. Perks and pitfalls of faithfulness in regular, self-explainable and domain invariant gnns, 2024. URL https://arxiv.org/abs/2406.15156

    work page Pith review arXiv 2024

  6. [6]

    and Albert, R

    Barab \'a si, A.-L. and Albert, R. Emergence of scaling in random networks. science, 286 0 (5439): 0 509--512, 1999

    1999

  7. [7]

    Entropy-based logic explanations of neural networks

    Barbiero, P., Ciravegna, G., Giannini, F., Li \'o , P., Gori, M., and Melacci, S. Entropy-based logic explanations of neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 6046--6054, 2022

    work page 2022

  8. [8]

    V., Monet, M., P \'e rez, J., Reutter, J., and Silva, J.-P

    Barcel \'o , P., Kostylev, E. V., Monet, M., P \'e rez, J., Reutter, J., and Silva, J.-P. The logical expressiveness of graph neural networks. In ICLR, 2020

    work page 2020

Show all 84 references
  1. [9]

    Graph neural networks use graphs when they shouldn't

    Bechler-Speicher, M., Amos, I., Gilad-Bachrach, R., and Globerson, A. Graph neural networks use graphs when they shouldn't. arXiv preprint arXiv:2309.04332, 2023

    2023 arXiv

  2. [10]

    The intelligible and effective graph neural additive network

    Bechler-Speicher, M., Globerson, A., and Gilad-Bachrach, R. The intelligible and effective graph neural additive network. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. URL https://openreview.net/forum?id=SKY1ScUTwA

    2024

  3. [11]

    Causal Explanations and XAI

    Beckers, S. Causal Explanations and XAI . In Conference on Causal Learning and Reasoning, pp.\ 90--109, 2022

    2022

  4. [12]

    and von Luxburg, U

    Bhattacharjee, R. and von Luxburg, U. Auditing local explanations is hard. arXiv preprint arXiv:2407.13281, 2024

    2024 arXiv

  5. [13]

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

    Chen, Y., Zhang, Y., Bian, Y., Yang, H., Kaili, M., Xie, B., Liu, T., Han, B., and Cheng, J. Learning causally invariant representations for out-of-distribution generalization on graphs. Advances in Neural Information Processing Systems, 35: 0 22131--22148, 2022

    2022

  6. [14]

    How interpretable are interpretable graph neural networks? In Forty-first International Conference on Machine Learning, 2024

    Chen, Y., Bian, Y., Han, B., and Cheng, J. How interpretable are interpretable graph neural networks? In Forty-first International Conference on Machine Learning, 2024

    2024

  7. [15]

    How faithful are self-explainable gnns? arXiv preprint arXiv:2308.15096, 2023

    Christiansen, M., Villadsen, L., Zhong, Z., Teso, S., and Mottin, D. How faithful are self-explainable gnns? arXiv preprint arXiv:2308.15096, 2023

    2023 arXiv

  8. [16]

    and Wang, S

    Dai, E. and Wang, S. Towards self-explainable graph neural network. In Proceedings of the 30th ACM International Conference on Information & Knowledge Management, pp.\ 302--311, 2021

    2021

  9. [17]

    and Wang, S

    Dai, E. and Wang, S. Towards prototype-based self-explainable graph neural network. arXiv preprint arXiv:2210.01974, 2022

    2022 arXiv

  10. [18]

    and Hirth, A

    Darwiche, A. and Hirth, A. On the (complete) reasons behind decisions. Journal of Logic, Language and Information, 32 0 (1): 0 63--88, 2023

    2023

  11. [19]

    K., Lopez de Compadre, R

    Debnath, A. K., Lopez de Compadre, R. L., Debnath, G., Shusterman, A. J., and Hansch, C. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds. correlation with molecular orbital energies and hydrophobicity. Journal of medicinal chemistry, 34...

    1991

  12. [20]

    and Shen, Y

    Deng, J. and Shen, Y. Self-interpretable graph learning with sufficient and necessary explanations. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pp.\ 11749--11756, 2024

    2024

  13. [21]

    On random graphs i

    Erdos, P., R \'e nyi, A., et al. On random graphs i. Publ. math. debrecen, 6 0 (290-297): 0 18, 1959

    1959

  14. [22]

    A self-explainable heterogeneous gnn for relational deep learning

    Ferrini, F., Longa, A., Passerini, A., and Jaeger, M. A self-explainable heterogeneous gnn for relational deep learning. arXiv preprint arXiv:2412.00521, 2024

    2024 arXiv

  15. [23]

    and Lenssen, J

    Fey, M. and Lenssen, J. E. Fast graph representation learning with pytorch geometric. arXiv preprint arXiv:1903.02428, 2019

    1903 arXiv

  16. [24]

    Xai and bias of deep graph networks

    Fontanesi, M., Micheli, A., and Podda, M. Xai and bias of deep graph networks. ESANN, 2024

    2024

  17. [25]

    C., Li \'o , P., and Barbiero, P

    Giannini, F., Fioravanti, S., Keskin, O., Lupidi, A., Magister, L. C., Li \'o , P., and Barbiero, P. Interpretable graph networks formulate universal algebra conjectures. Advances in Neural Information Processing Systems, 36, 2024

    2024

  18. [26]

    V., Gonzalez, G., and Agarwal, C

    Giunchiglia, V., Shukla, C. V., Gonzalez, G., and Agarwal, C. Towards training GNN s using explanation directed message passing. In The First Learning on Graphs Conference, 2022. URL https://openreview.net/forum?id=_nlbNbawXDi

    2022

  19. [27]

    Anti-symmetric dgn: a stable architecture for deep graph networks

    Gravina, A., Bacciu, D., and Gallicchio, C. Anti-symmetric dgn: a stable architecture for deep graph networks. In The Eleventh International Conference on Learning Representations, 2022

    2022

  20. [28]

    The logic of graph neural networks

    Grohe, M. The logic of graph neural networks. In 36th Annual ACM/IEEE Symposium on Logic in Computer Science (LICS), pp.\ 1--17, 2021

    2021

  21. [29]

    GOOD : A graph out-of-distribution benchmark

    Gui, S., Li, X., Wang, L., and Ji, S. GOOD : A graph out-of-distribution benchmark. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. URL https://openreview.net/forum?id=8hHg-zs_p-h

    2022

  22. [30]

    Joint learning of label and environment causal independence for graph out-of-distribution generalization

    Gui, S., Liu, M., Li, X., Luo, Y., and Ji, S. Joint learning of label and environment causal independence for graph out-of-distribution generalization. arXiv preprint arXiv:2306.01103, 2023

    2023 arXiv

  23. [31]

    Addressing leakage in concept bottleneck models

    Havasi, M., Parbhoo, S., and Doshi-Velez, F. Addressing leakage in concept bottleneck models. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K. (eds.), Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?id=tglniD_fn9

    2022

  24. [32]

    Sat-based formula simplification

    Ignatiev, A., Previti, A., and Marques-Silva, J. Sat-based formula simplification. In Theory and Applications of Satisfiability Testing -- SAT 2015, pp.\ 287--298. Springer International Publishing, 2015

    2015

  25. [33]

    On relating explanations and adversarial examples

    Ignatiev, A., Narodytska, N., and Marques-Silva, J. On relating explanations and adversarial examples. NeurIPS, 2019

    2019

  26. [34]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Ioffe, S. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015

    2015 arXiv

  27. [35]

    Categorical reparameterization with gumbel-softmax

    Jang, E., Gu, S., and Poole, B. Categorical reparameterization with gumbel-softmax. arXiv preprint arXiv:1611.01144, 2016

    2016 arXiv

  28. [36]

    Multi-objective molecule generation using interpretable substructures

    Jin, W., Barzilay, R., and Jaakkola, T. Multi-objective molecule generation using interpretable substructures. In International conference on machine learning, pp.\ 4849--4859. PMLR, 2020

    2020

  29. [37]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In International Conference on Learning Representations, 2015

    2015

  30. [38]

    P., Mesiar, R., and Pap, E

    Klement, E. P., Mesiar, R., and Pap, E. Triangular norms, volume 8. Springer Science & Business Media, 2013

    2013

  31. [39]

    W., and Amer, M

    Knyazev, B., Taylor, G. W., and Amer, M. Understanding attention and generalization in graph neural networks. Advances in neural information processing systems, 32, 2019

    2019

  32. [40]

    and Heindorf, S

    K \"o hler, D. and Heindorf, S. Utilizing description logics for global explanations of heterogeneous graph neural networks. arXiv preprint arXiv:2405.12654, 2024

    2024 arXiv

  33. [41]

    Estimating mutual information

    Kraskov, A., St \"o gbauer, H., and Grassberger, P. Estimating mutual information. Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 69 0 (6): 0 066138, 2004

    2004

  34. [42]

    Explainable graph neural networks under fire, 2024

    Li, Z., Geisler, S., Wang, Y., Günnemann, S., and van Leeuwen, M. Explainable graph neural networks under fire, 2024. URL https://arxiv.org/abs/2406.06417

    2024 arXiv

  35. [43]

    J., Bulusu, K

    Lin, C., Sun, G. J., Bulusu, K. C., Dry, J. R., and Hernandez, M. Graph neural networks including sparse interpretability. arXiv preprint arXiv:2007.00119, 2020

    2007 arXiv

  36. [44]

    Explaining the explainers in graph neural networks: a comparative study

    Longa, A., Azzolin, S., Santin, G., Cencetti, G., Li \`o , P., Lepri, B., and Passerini, A. Explaining the explainers in graph neural networks: a comparative study. ACM Computing Surveys, 2024

    2024

  37. [45]

    Parameterized explainer for graph neural network

    Luo, D., Cheng, W., Xu, D., Yu, W., Zong, B., Chen, H., and Zhang, X. Parameterized explainer for graph neural network. Advances in neural information processing systems, 33: 0 19620--19631, 2020

    2020

  38. [46]

    Do concept bottleneck models learn as intended? arXiv preprint arXiv:2105.04289, 2021

    Margeloiu, A., Ashman, M., Bhatt, U., Chen, Y., Jamnik, M., and Weller, A. Do concept bottleneck models learn as intended? arXiv preprint arXiv:2105.04289, 2021

    2021 arXiv

  39. [47]

    Logic-based explainability in machine learning

    Marques-Silva, J. Logic-based explainability in machine learning. In Reasoning Web. Causality, Explanations and Declarative Knowledge: 18th International Summer School 2022, Berlin, Germany, September 27--30, 2022, Tutorial Lectures, pp.\ 24--104. Springer, 2023

    2022

  40. [48]

    and Ignatiev, A

    Marques-Silva, J. and Ignatiev, A. Delivering trustworthy ai through formal xai. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 12342--12350, 2022

    2022

  41. [49]

    C., Ignatiev, A., and Narodytska, N

    Marques-Silva, J., Gerspacher, T., Cooper, M. C., Ignatiev, A., and Narodytska, N. Explaining naive bayes and other linear classifiers with polynomial time and delay. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, 2020. ...

    2020

  42. [50]

    F., Teixeira, A

    Martins, I. F., Teixeira, A. L., Pinheiro, L., and Falcao, A. O. A bayesian approach to in silico blood-brain barrier penetration modeling. Journal of chemical information and modeling, 52 0 (6): 0 1686--1697, 2012

    2012

  43. [51]

    and Stratos, K

    McAllester, D. and Stratos, K. Formal limitations on the measurement of mutual information. In International Conference on Artificial Intelligence and Statistics, pp.\ 875--884. PMLR, 2020

    2020

  44. [52]

    Interpretable and generalizable graph learning via stochastic attention mechanism

    Miao, S., Liu, M., and Li, P. Interpretable and generalizable graph learning via stochastic attention mechanism. In International Conference on Machine Learning, pp.\ 15524--15543. PMLR, 2022 a

    2022

  45. [53]

    Interpretable geometric deep learning via learnable randomness injection

    Miao, S., Luo, Y., Liu, M., and Li, P. Interpretable geometric deep learning via learnable randomness injection. In The Eleventh International Conference on Learning Representations, 2022 b

    2022

  46. [54]

    M., Bause, F., Kersting, K., Mutzel, P., and Neumann, M

    Morris, C., Kriege, N. M., Bause, F., Kersting, K., Mutzel, P., and Neumann, M. Tudataset: A collection of benchmark datasets for learning with graphs. arXiv preprint arXiv:2007.08663, 2020

    2007 arXiv

  47. [55]

    Graphchef: Learning the recipe of your dataset

    M \"u ller, P., Faber, L., Martinkus, K., and Wattenhofer, R. Graphchef: Learning the recipe of your dataset. In ICML 3rd Workshop on Interpretable Machine Learning in Healthcare (IMLH), 2023. URL https://openreview.net/forum?id=ZgYZH5PFEg

    2023

  48. [56]

    Automatic differentiation in pytorch

    Paszke, A., Gross, S., Chintala, S., Chanan, G., Yang, E., DeVito, Z., Lin, Z., Desmaison, A., Antiga, L., and Lerer, A. Automatic differentiation in pytorch. In NIPS-W, 2017

    2017

  49. [57]

    Logical distillation of graph neural networks

    Pluska, A., Welke, P., G \"a rtner, T., and Malhotra, S. Logical distillation of graph neural networks. In ICML 2024 Workshop on Mechanistic Interpretability, 2024

    2024

  50. [58]

    Prototype-based interpretable graph neural networks

    Ragno, A., La Rosa, B., and Capobianco, R. Prototype-based interpretable graph neural networks. IEEE Transactions on Artificial Intelligence, 2022

    2022

  51. [59]

    and Bunke, H

    Riesen, K. and Bunke, H. Iam graph database repository for graph based pattern recognition and machine learning. In Structural, Syntactic, and Statistical Pattern Recognition: Joint IAPR International Workshop, SSPR & SPR 2008, Orlando, USA, December 4-6, 2008. Proceedings, pp...

    2008

  52. [60]

    Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead

    Rudin, C. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nature Machine Intelligence, 1 0 (5): 0 206--215, 2019

    2019

  53. [61]

    C., Hagenbuchner, M., and Monfardini, G

    Scarselli, F., Gori, M., Tsoi, A. C., Hagenbuchner, M., and Monfardini, G. The graph neural network model. IEEE transactions on neural networks, 20 0 (1): 0 61--80, 2008

    2008

  54. [62]

    and Niepert, M

    Serra, G. and Niepert, M. Learning to explain graph neural networks. arXiv preprint arXiv:2209.14402, 2022

    2022 arXiv

  55. [63]

    Glocalx-from local to global explanations of black box ai models

    Setzu, M., Guidotti, R., Monreale, A., Turini, F., Pedreschi, D., and Giannotti, F. Glocalx-from local to global explanations of black box ai models. Artificial Intelligence, 294: 0 103457, 2021

    2021

  56. [64]

    A symbolic approach to explaining bayesian network classifiers

    Shih, A., Choi, A., and Darwiche, A. A symbolic approach to explaining bayesian network classifiers. arXiv preprint arXiv:1805.03364, 2018

    2018 arXiv

  57. [65]

    Axiomatic attribution for deep networks

    Sundararajan, M., Taly, A., and Yan, Q. Axiomatic attribution for deep networks. In International conference on machine learning, pp.\ 3319--3328. PMLR, 2017

    2017

  58. [66]

    A., Poda, G., and Tetko, I

    Sushko, I., Salmina, E., Potemkin, V. A., Poda, G., and Tetko, I. V. Toxalerts: A web server of structural alerts for toxic chemicals and compounds with potential adverse reactions. Journal of Chemical Information and Modeling, 52 0 (8): 0 2310--2316, 2012. doi:10.1021/ci300245q

    2012 doi

  59. [67]

    Learning and evaluating graph neural network explanations based on counterfactual and factual reasoning

    Tan, J., Geng, S., Fu, Z., Ge, Y., Xu, S., Li, Y., and Zhang, Y. Learning and evaluating graph neural network explanations based on counterfactual and factual reasoning. In Proceedings of the ACM Web Conference 2022, WWW '22, pp.\ 1018–1027, New York, NY, USA, 2022. Associatio...

    2022

  60. [68]

    Leveraging explanations in interactive machine learning: An overview

    Teso, S., Alkan, \"O ., Stammer, W., and Daly, E. Leveraging explanations in interactive machine learning: An overview. Frontiers in Artificial Intelligence, 2023

    2023

  61. [69]

    C., and Bialek, W

    Tishby, N., Pereira, F. C., and Bialek, W. The information bottleneck method. ArXiv, physics/0004057, 2000. URL https://api.semanticscholar.org/CorpusID:8936496

    2000 arXiv

  62. [70]

    Probabilistic sufficient explanations

    Wang, E., Khosravi, P., and Van den Broeck, G. Probabilistic sufficient explanations. 2021

    2021

  63. [71]

    Weisstein, E. W. Edge cover, 2025. URL https://mathworld.wolfram.com/EdgeCover.html

    2025

  64. [72]

    J., Valeri, J

    Wong, F., Zheng, E. J., Valeri, J. A., Donghia, N. M., Anahtar, M. N., Omori, S., Li, A., Cubillos-Ruiz, A., Krishnan, A., Jin, W., et al. Discovery of a structural class of antibiotics with explainable deep learning. Nature, 626 0 (7997): 0 177--185, 2024

    2024

  65. [73]

    Graph information bottleneck

    Wu, T., Ren, H., Li, P., and Leskovec, J. Graph information bottleneck. Advances in Neural Information Processing Systems, 33: 0 20437--20448, 2020 a

    2020

  66. [74]

    Discovering invariant rationales for graph neural networks

    Wu, Y.-X., Wang, X., Zhang, A., He, X., and Chua, T.-S. Discovering invariant rationales for graph neural networks. arXiv preprint arXiv:2201.12872, 2022

    2022 arXiv

  67. [75]

    N., Gomes, J., Geniesse, C., Pappu, A

    Wu, Z., Ramsundar, B., Feinberg, E. N., Gomes, J., Geniesse, C., Pappu, A. S., Leswing, K., and Pande, V. Moleculenet: a benchmark for molecular machine learning. Chemical science, 9 0 (2): 0 513--530, 2018

    2018

  68. [76]

    Wu, Z., Pan, S., Chen, F., Long, G., Zhang, C., and Philip, S. Y. A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 2020 b

    2020

  69. [77]

    How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018

    Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826, 2018

    2018 arXiv

  70. [78]

    Gnnexplainer: Generating explanations for graph neural networks

    Ying, Z., Bourgeois, D., You, J., Zitnik, M., and Leskovec, J. Gnnexplainer: Generating explanations for graph neural networks. Advances in neural information processing systems, 32, 2019

    2019

  71. [79]

    Graph information bottleneck for subgraph recognition

    Yu, J., Xu, T., Rong, Y., Bian, Y., Huang, J., and He, R. Graph information bottleneck for subgraph recognition. In International Conference on Learning Representations, 2020

    2020

  72. [80]

    Improving subgraph recognition with variational graph information bottleneck

    Yu, J., Cao, J., and He, R. Improving subgraph recognition with variational graph information bottleneck. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 19396--19405, 2022

    2022

  73. [81]

    Yu, J., Ignatiev, A., and Stuckey, P. J. On formal feature attribution and its approximation. arXiv preprint arXiv:2307.03380, 2023

    2023 arXiv

  74. [82]

    On explainability of graph neural networks via subgraph explorations

    Yuan, H., Yu, H., Wang, J., Li, K., and Ji, S. On explainability of graph neural networks via subgraph explorations. In International conference on machine learning, pp.\ 12241--12252. PMLR, 2021

    2021

  75. [83]

    Explainability in graph neural networks: A taxonomic survey

    Yuan, H., Yu, H., Gui, S., and Ji, S. Explainability in graph neural networks: A taxonomic survey. IEEE transactions on pattern analysis and machine intelligence, 45 0 (5): 0 5782--5799, 2022

    2022

  76. [84]

    Protgnn: Towards self-explaining graph neural networks

    Zhang, Z., Liu, Q., Wang, H., Lu, C., and Lee, C. Protgnn: Towards self-explaining graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pp.\ 9127--9135, 2022

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.