Pith. sign in

REVIEW 4 major objections 5 minor 27 references

Graph Inverse Style Transfer for Counterfactual Explainability

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Graph counterfactuals improve when generated by overshooting the decision boundary and then backtracking via spectral style transfer: +7.6% validity and +45.5% fidelity over the second-best baselines on 8 benchmarks.

desk verdict The backtracking idea is genuinely new and the writing is honest, but GIST's content loss is pinned to a random other-class graph, not the input, so the generated graphs are not minimal counterfactuals of anything. read the letter →

arxiv 2505.17542 v2 pith:EUTTEUNE submitted 2025-05-23 cs.LG

Bardh Prenkaj , Efstratios Zaradoukas , Gjergji Kasneci This is my paper · ORCID
classification cs.LG MSC 05C5068T07
keywords graphcounterfactualexplainabilityspectralstyletransfernormalizedLaplacianbacktrackingneuralnetworksvaliditygapGumbel-Softmaxedgesampling
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

The paper tries to establish that graph counterfactual explanations are better produced backwards than forwards. Instead of nudging an input graph until the underlying model changes its verdict, GIST first overshoots the decision boundary to a different-class graph and then learns a reverse transformation that pulls the candidate back toward the input's global spectral structure while keeping the overshoot graph's local content. The authors argue this makes the overshooting problem controllable through a single interpolation factor and produces counterfactuals that are more often valid and more faithful to the true class distribution. If correct, the claim matters because it reframes a search problem that tends to over-edit or destroy structure as a style-transfer problem with explicit control over the validity-proximity trade-off.

What carries the argument

The load-bearing object is the convex combination of normalized Laplacians, $\tilde{L}(G^*) = \alpha \tilde{L}(G_\varepsilon) + (1-\alpha)\tilde{L}(G)$, together with the spectral style loss it justifies: the $\ell^1$ distance between the sorted eigenvalues of the input and the counterfactual. When the two Laplacians commute and share the same size, the eigenvalues of $G^*$ are exactly the convex combination of the endpoints' eigenvalues (Lemma 4.3), which yields connectivity preservation (Theorem 4.4), a between-bounds spectral gap (Theorem 4.5), and the Frobenius identity $\|L(G_\varepsilon) - L(G^*)\|_F = (1-\alpha)\|L(G_\varepsilon) - L(G)\|_F$ (Corollary 4.6). The learned backtracking network — transformer convolution layers on the overshoot graph, an MLP edge-probability head, and Gumbel-Softmax edge sampling — is the mechanism that realizes this interpolation on real graphs.

What would settle it

Set $\alpha = 0.5$ instead of 0.9 and test the paper's central in-between prediction, $\|L(G_\varepsilon) - L(G^*)\|_F \approx (1-\alpha)\|L(G_\varepsilon) - L(G)\|_F$: on AIDS the paper's own experiments report an error of 0.537 at $\alpha = 0.5$, against errors of order $10^{-3}$ at $\alpha = 0.9$, so the interpolation story must be checked across the whole $\alpha$-range rather than at the extreme where it was validated. A second check: on a multi-class benchmark, count how often the produced counterfactual lands in a third class rather than the overshoot class — the paper explicitly disclaims any guarantee that $\Phi(G^*) = \Phi(G_\varepsilon)$.

Watch

Extended reading notes

Core claim

On the paper's own terms, GIST's central claim is that a counterfactual graph $G^*$ can be obtained as a spectral interpolation between the input graph $G$ and a different-class overshoot graph $G_\varepsilon$: the normalized Laplacian of $G^*$ is modeled as $\tilde{L}(G^*) = \alpha \tilde{L}(G_\varepsilon) + (1-\alpha)\tilde{L}(G)$, so its eigenvalues, spectral gap, and Frobenius deviation from $G_\varepsilon$ all inherit bounds from the two endpoint graphs. The learning objective realizes this interpolation by minimizing $\alpha$ times a content-preservation term (node-feature and edge reconstruction against $G_\varepsilon$) plus $(1-\alpha)$ times a style term, namely the $\ell^1$ distance between the sorted normalized-Laplacian eigenvalues of $G$ and $G^*$. Empirically the paper reports that this backtracking procedure beats forward-perturbation baselines on the benchmarks used, with the largest validity gains on multi-class tasks and a reported +45.5% improvement in fidelity.

Load-bearing premise

The theory assumes that Laplacians of graphs with different node counts can be compared by padding adjacency and degree matrices to a common size, so that sorted eigenvalues can be matched entry-by-entry; the exact interpolation theorems additionally require equal-size, commuting Laplacians.

Editorial extensions

If this is right

  • If GIST is right, forward perturbation is not the only workable paradigm for graph counterfactuals: crossing the boundary first and then refining gives higher validity on 4 of 8 benchmarks, parity on the rest, and higher fidelity almost everywhere.
  • The interpolation factor $\alpha$ becomes a direct control dial: moving $\alpha$ toward 1 pushes counterfactuals closer to the overshoot graph and raises validity, at the cost of larger spectral distance from the input (Figure 3).
  • In the idealized equal-size setting, produced counterfactuals inherit connectivity and a spectral gap bounded by the two endpoint graphs, so structural coherence is guaranteed rather than tuned.
  • Spectral distance (sorted-eigenvalue $\ell^1$ distance) is proposed as a more principled proximity measure than Graph Edit Distance, since edge edits that do not move the spectrum are treated as structurally neutral.
  • The framework requires no paired counterfactual data: the overshoot graph is any different-class graph from the dataset, so the method applies wherever an oracle and a dataset exist.

Reading between the lines

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

  • A testable corollary the paper does not draw: if spectral alignment is the effective mechanism, GIST-type methods should dominate on benchmarks where classes differ by global connectivity (trees versus cycles) and lose their edge on motif-localized classes, where local edits suffice.
  • The exact theory needs equal-size commuting Laplacians; since real benchmarks vary in size, the sorted-eigenvalue $\ell^1$ comparison rests on zero padding. Swapping in a Wasserstein distance on spectra — which the paper names as future work — would show whether the interpolation story survives variable-size graphs.
  • The random overshoot selection leaves performance hostage to which different-class graph is drawn; the paper sketches but does not run a proximity-conditioned variant (Equation 27). If that variant materially improves convergence, much of the reported gain may come from spectral alignment rather than the backtracking mechanism itself.
  • The reported degradation at $\alpha = 0.5$ suggests practitioners should treat the in-between interpolation guarantee as valid only in the content-dominant regime.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes GIST (Graph Inverse Style Transfer), a method for graph counterfactual explainability that first overshoots the decision boundary of a black-box oracle by selecting a random opposite-class graph Gε and then learns a backtracking mechanism that generates a graph G* whose normalized Laplacian spectrum is aligned with the input G while its local content is aligned with Gε. The method is formalized through a loss function (Eq. 12) that combines a content reconstruction term with respect to Gε and a spectral style term with respect to G. The authors provide theoretical results for convex combinations of Laplacians (spectral interpolation, connectivity preservation, spectral-gap bounds, and a Frobenius-norm identity) and evaluate GIST on eight graph classification benchmarks against CF-GNNExp., CF2, CLEAR, and RSGG-CE, reporting improved validity and fidelity.

Significance. The idea of re-framing graph counterfactual generation as a spectral style-transfer backtracking process is novel and could open a useful new direction if the generated graphs were faithful minimal edits of the input. The paper also provides an unusually detailed appendix, a public code link, and extensive ablations, which are commendable. However, the central claim is not supported: the loss function preserves content of a randomly selected opposite-class graph rather than of the input, so the generated graphs are not minimal counterfactual perturbations of G under the paper's own definition (Eq. 3). The theoretical guarantees apply only to equal-size commuting Laplacians, while the benchmarks are variable-size and require padding, a limitation acknowledged in Section 5.3. The reported improvements in validity and fidelity may therefore reflect copying target-class structure rather than explaining the specific input. Given that this issue is load-bearing and cannot be repaired without redefining the objective, the paper's main contribution as a counterfactual explainer is not established.

major comments (4)
  1. [Section 4.2.2, Eq. (12)] The optimization objective minimizes content reconstruction against the randomly selected opposite-class graph Gε, not against the input G. The content term is ||X* - Xε||1 + BCE(A*, Aε), and with α=0.9 it dominates the loss, pulling G* toward Gε. This is confirmed by Figure 3, which shows GED(G,G*) increasing as α→1, and by Table 5, where GIST's GED between G and G* on AIDS is 52.0 versus 2.3 for iRand. This contradicts the counterfactual definition in Eq. (3) and the abstract's claim of 'minimal changes' to the input. The generated graphs are interpolations between G's spectrum and Gε's content, so the reported validity/fidelity gains in Tables 1-2 may measure how well GIST copies target-class structure rather than how faithfully it explains the specific input.
  2. [Section 4.1, Definition 4.2 vs. Section 5.3] The theoretical results require L~(G) and L~(Gε) to be real, symmetric, commuting matrices of the same dimension. The benchmark graphs have different numbers of nodes (Table 3), and Section 5.3 explicitly states that the zeros in Figures 4-6 are due to padding adjacency and degree matrices to a common dimension. Padding appends zero eigenvalues and alters the spectrum, and the convex combination in Eq. (7) is undefined for matrices of different sizes. Consequently, Lemma 4.3, Theorems 4.4-4.5, and Corollary 4.6 do not apply to the reported experiments. Moreover, the spectral 'validation' in Figures 4-6 checks quantities that are directly minimized by Lstyle in Eq. (12), so it is an optimization sanity check rather than evidence of semantic or structural preservation.
  3. [Tables 1, 2, and 5] The experimental results are internally inconsistent. For AIDS, Table 1 reports GIST validity of 0.969 and Table 2 reports fidelity of 0.957, while Table 5 reports validity of 0.940 and fidelity of 0.928 under the same 5-fold cross-validation protocol. No standard deviations or confidence intervals are reported for any of the tables, so the headline improvements of +7.6% in validity and +45.5% in fidelity cannot be assessed for statistical significance. At minimum, the AIDS discrepancy must be explained, and all aggregate claims need variance estimates.
  4. [Section 6, Limitations] The paper concedes that there is no guarantee of class maintenance during backtracking in multiclass scenarios. This is not a minor caveat: for the multiclass datasets ENZYMES and COLORS-3, a generated graph that changes the oracle's prediction to any of the other classes counts as valid in Table 1, even if it does not land in the intended counterfactual class. Without per-class validity results or a constraint that Φ(G*) equals the target class, the multiclass validity numbers conflate arbitrary class flips with meaningful counterfactual explanations.
minor comments (5)
  1. [Section 4.1] The sentence 'if G and Gε are connected graphs, then also Gε is connected' should read 'then also G* is connected'; as written it is a tautology and inconsistent with the surrounding discussion.
  2. [Section 5.3] The text says 'we expect that the GED between G and Gε decreases when α→1 (see the red curve)', but Figure 3's red curve is described as the distance between G* and Gε; this should be GED(G*,Gε).
  3. [Section 5.2] The abbreviation 'RSSG-CE' appears once; it should be 'RSGG-CE' for consistency with the rest of the paper.
  4. [Section 5.1] The word 'explainations' should be 'explanations'.
  5. [Tables 1-2] Since all results are averages over 5-fold cross-validation, reporting standard deviations or individual fold results would substantially improve interpretability, especially for the claimed average gains.

Circularity Check

2 steps flagged · score 4.0 of 10

Spectral 'validations' restate the training objective, and the 'interpolation' claim is a paraphrase of the loss; benchmark comparisons remain independent.

  1. fitted input called prediction [Section 5.3, Eq. (12), Definition 4.2, Figure 4]
    "GIST finds counterfactuals whose eigenvalues are closely aligned to the optimal convex combination of Equation (8). ... Here, we illustrate the eigenvalues of G∗ as produced by GIST, and the true combined spectra with α = 0.9."

    The training loss (Eq. 12) is α[||X*−Xε||1 + BCE(A*,Aε)] + (1−α)Σ|λi(L̃(G))−λi(L̃(G*))|. The 'optimal' spectrum used in Figure 4 is αL̃(Gε)+(1−α)L̃(G), which is the same α-weighted blend of the same two targets (Gε content, G spectrum) that the loss minimizes. A model trained to be close to Gε with weight 0.9 and to G's spectrum with weight 0.1 will necessarily land near that convex combination, so the reported 2×10−3 error is a display of the objective, not an independent confirmation. Corollary 4.6's Frobenius identity is an algebraic restatement of Definition 4.2, and Figure 6 reproduces it through the same loss, so these spectral validations reduce to the training objective by construction.

  2. self definitional [Definition 4.1, Eq. (6), Eq. (12), Figure 3]
    "Content preservation – the local structure of G∗ resembles Gε, i.e., min ||X*−Xε||1 + BCE(A*,Aε). ... As α → 1, the distance between G and G∗ becomes larger; the distance between G∗ and Gε decreases."

    The paper defines 'content faithfulness' as reconstruction of Gε, not of the input G. With α=0.9, the content term dominates and the only term involving G is the 0.1-weighted spectral L1 loss. Hence the produced G* is by construction close to the randomly chosen opposite-class graph Gε and only weakly tied to G. The abstract's claim that GIST produces 'valid counterfactuals as interpolations between the input style and counterfactual content' is therefore a paraphrase of the loss itself, while the minimal-change requirement of Eq. (3) is not actually optimized. The reported validity is largely inherited from Gε's opposite-class label rather than being established as a minimal edit of the explained input.

full rationale

The direct comparison against forward baselines (Tables 1-2) is a genuine empirical evaluation and is not circular. The convex-combination algebra (Lemmas 4.3, Theorems 4.4-4.5, Corollary 4.6) is a standard, independently provable mathematical fact, though its applicability to the learned G* is not established. The partial circularity lies in the paper's spectral 'validation': Figures 4-6 measure exactly the trade-off that Eq. (12) is trained to optimize, using the same α, so the small reported errors restate the objective rather than test the theory. Additionally, the central narrative of 'interpolation between input style and counterfactual content' is definitionally tied to Gε because content preservation is defined as reconstruction of Gε. These issues make the explanatory claim partially circular, but the benchmark results retain independent content, so a score of 4 is appropriate rather than 6 or higher.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The GIST framework rests on the modeling assumption that a graph's normalized Laplacian spectrum is its 'style' and that a random opposite-class graph's content is appropriate to preserve. The empirical comparison treats α as fixed at 0.9; the spectral results assume same-size commuting Laplacians and are evaluated with padding. No new physical or mathematical entities are introduced.

free parameters (1)
  • alpha (interpolation factor) = 0.9
    Controls the balance between content (Gε) and style (G) in Eq. 12; set to 0.9 across all datasets 'to encourage higher validity' (Appendix C). The main results and all spectral validation figures depend on this hand-chosen value; α=0.5 degrades performance (Appendix G.4).
assumptions (4)
  • ad hoc to paper Normalized Laplacians of G and Gε can be compared/aligned by padding to a common dimension
    Used in Eq. 5, Eq. 12, and Figures 4-6. The paper notes zeros due to padding and defers Wasserstein distance to future work. For graphs of different sizes, eigenvalue sums are not well-defined without padding.
  • ad hoc to paper Eigenvalues of a graph's normalized Laplacian capture its semantic 'style' and are a valid proxy for structural meaning in counterfactuals
    Core modeling assumption in Section 2 and Definition 4.1; no independent evidence that spectral similarity preserves semantic validity for molecules or proteins.
  • ad hoc to paper The chosen random Gε from the dataset is a suitable 'overshot' starting point; preserving its content yields a meaningful counterfactual
    Equation (26) selects the first graph of a different class from a shuffled dataset; Lcont forces G* to reconstruct Xε and Aε, so the counterfactual content is inherited from an arbitrary graph rather than the input G.
  • domain assumption The Laplacians of G and Gε commute (or the non-commuting bound suffices)
    Definition 4.2 and Lemma 4.3 assume real symmetric commuting matrices for exact eigenvalue interpolation; Appendix A.1.2 falls back to Weyl bounds when they do not commute. No evidence that real benchmark pairs commute.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph Inverse Style Transfer for Counterfactual Explainability." pith.science (2026). https://pith.science/paper/EUTTEUNE

@misc{pith2026250517542,
  author       = {Pith},
  title        = {Pith review of: Graph Inverse Style Transfer for Counterfactual Explainability},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EUTTEUNE}},
  note         = {Machine review of arXiv:2505.17542}
}
read the original abstract

Counterfactual explainability seeks to uncover model decisions by identifying minimal changes to the input that alter the predicted outcome. This task becomes particularly challenging for graph data due to preserving structural integrity and semantic meaning. Unlike prior approaches that rely on forward perturbation mechanisms, we introduce Graph Inverse Style Transfer (GIST), the first framework to re-imagine graph counterfactual generation as a backtracking process, leveraging spectral style transfer. By aligning the global structure with the original input spectrum and preserving local content faithfulness, GIST produces valid counterfactuals as interpolations between the input style and counterfactual content. Tested on 8 binary and multi-class graph classification benchmarks, GIST achieves a remarkable +7.6% improvement in the validity of produced counterfactuals and significant gains (+45.5%) in faithfully explaining the true class distribution. Additionally, GIST's backtracking mechanism effectively mitigates overshooting the underlying predictor's decision boundary, minimizing the spectral differences between the input and the counterfactuals. These results challenge traditional forward perturbation methods, offering a novel perspective that advances graph explainability.

Figures

Figures reproduced from arXiv: 2505.17542 by the authors.

Figure 1
Figure 1. The “gist” of GIST. Starting from an initial graph G ∈ G \ G′ , we overshoot to the other side of Φ’s boundary on a randomly chosen graph G ε ∈ G′ . We then learn a reverse pro￾cess to backtrack from G ε to a never-seen-before graph G ∗ ∈ G′ while imitating the global structure or style (Lstyle) of G and main￾taining faithful local structure and preserve content (Lcont) with G ε . Notice how Lstyle pulls the generat… view at source ↗
Figure 2
Figure 2. Overview of GIST’s forward learning pass and backtracking optimization. Given a graph G = (X, A), we use Φ to overshoot to G ε = (X ε , Aε ) s.t. Φ(G) ̸= Φ(G ε ). We feed G ε through transformer convolution layers that output node embeddings X ∗ . For each embedding, we create edge pairs and feed them to an MLP which estimates probabilities pi,j . We can sample these probabilities to compose a new adjacency matrix A… view at source ↗
Figure 3
Figure 3. As α → 1, the distance between G and G∗ becomes larger; the distance between G∗ and Gε decreases. We show the relationship between the interpolation factor α, the GED, and the validity of G ∗ as guided by the oracle Φ on AIDS. Notice how the validity of GIST increases when the distance between G ∗ and G increases. 5.3. Ablation studies The interpolation factor α controls the trade-off between the distances of G, G∗ … view at source ↗
Figures from the paper (10 more)
Figure 5
Figure 5. Figure 5: GIST finds G∗ whose gap satisfies Theorem 4.5 with only 1.3 × 10−2 difference against the optimal spectral gap. (left) We show the spectral gaps of G, G ε and the found G ∗ according to α = 0.9 on 100 random samples of MSRC21. We also show the optimal counterfactual to…
Figure 4
Figure 4. Figure 4: GIST’s backtracking mechanism finds G∗ with under 2.051 × 10−3 error vs. the optimal counterfactual in terms of spectra convex combination in Equation (8). We show the eigenvalues of 10 random (G, Gε ) pairs on AIDS, and the finding of G ∗ according to GIST with α = 0.…
Figure 7
Figure 7. Figure 7: Taken from (Neumann et al., 2016). The left-most RGB image is represented by a graph of superpixels (middle) with semantic labels b building, c car, v void, and ? unlabeled. (right) Point clouds of household objects represented by labeled 4-nearest-neighbor graphs with…
Figure 8
Figure 8. Figure 8: GIST is the most effective and reliable explainer (in terms of validity), reporting low input structural changes. Comparison of GED across different explainers for generating counterfactuals on various datasets. GIST consistently achieves the best balance between valid…
Figure 9
Figure 9. Figure 9: Heatmaps showing the spectra of G ∗ , obtained as a convex combination of the spectra of G (tree) and G ε (cyclic graph) for 10 different pairs of G and G ε . The interpolation parameter α varies from 0.1 to 1, transitioning from the eigenvalues of G (α = 0) to those o…
Figure 10
Figure 10. Figure 10: We show the spectral gaps of G, G ε , and G ∗ according to α = 0.9 on 100 random samples of AIDS (up left), COLORS-3 (up right), ENZYMES (down left), and PROTEIN (down right). We also show the optimal counterfactual to assess the distance with the spectral gap of G ∗ …
Figure 11
Figure 11. Figure 11: We show the Frobenius norm differences between the Laplacians of G, G ε and G ∗ on AIDS (row=1, col=1), BZR (row=1, col=2), COLORS-3 (row=2, col=1), ENZYMES (row=2, col=2), and MSRC21 (row=3) for α = 0.9. kind of GCN (with any layer) and U2GNN (Nguyen et al., 2022) wi…
Figure 12
Figure 12. Figure 12: We show the eigenvalues of 100 randomly sampled instances on PROTEINS (α = 0.9). explainability research. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: We show the Frobenius norm differences between the Laplacians of G, G ε and G ∗ on AIDS for α = 0.5. 0 20 40 60 80 100 (G, G ) Pair 10 9 10 7 10 5 10 3 10 1 S p e c t r u m G a p ( ) (G), (G * ), and (G ), Gap Error = 0.011 (G) (G ) Optimal (G * ) GIST (G * ) 0 20 40 …
Figure 14
Figure 14. Figure 14: We show the spectral gaps of G, G ε , and G ∗ according to α = 0.5 on 100 random samples of MSRC21. We also show the optimal counterfactual to assess the distance with the spectral gap of G ∗ . Each subplot on the right illustrates the first eigenvalue of G, G ε , and…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 23 canonical work pages

  1. [1]

    Notice that GIST, before performing the backtracking mechanism, needs to overshoot the decision boundary of Φ given an input graph G

    into finding the counterfactuals G∗. Notice that GIST, before performing the backtracking mechanism, needs to overshoot the decision boundary of Φ given an input graph G. In this paper, given G, we rely on a simple overshooting algorithm which finds Gε as in Equation (26). k∗ = min k ∈ [1, n] | Φ(G) ̸= Φ(Gk) ∀Gk ∈ U ({G | G ∈ G), Gε := Gk∗ , (26) 15 Graph...

    work page 2022

  2. [3]

    Categorical reparametrization with gumble-softmax

    Jang, E., Gu, S., and Poole, B. Categorical reparametrization with gumble-softmax. In International Conference on Learning Representations (ICLR 2017). OpenReview. net,

    work page 2017

  3. [6]

    Coun- tergan: Generating counterfactuals for real-time recourse and interpretability using residual gans

    Nemirovsky, D., Thiebaut, N., Xu, Y ., and Gupta, A. Coun- tergan: Generating counterfactuals for real-time recourse and interpretability using residual gans. In Cussens, J. and Zhang, K. (eds.), Uncertainty in Artificial Intelligence, Proceedings of the Thirty-Eighth Conference on Uncer- tainty in Artificial Intelligence, UAI 2022, 1-5 August 2022, Eindh...

    work page 2022

  4. [7]

    Nodes are connected if the superpixels are adjacent, and each node can further be annotated with a semantic label

    The nodes of each graph are derived by over-segmenting the images using the quick shift algorithm,7 resulting in one graph among the superpixels of each image. Nodes are connected if the superpixels are adjacent, and each node can further be annotated with a semantic label. Imagining an image retrieval system, where users provide images with semantic info...

    work page 2005

  5. [10]

    doi: 10.1145/3618105

    ISSN 0360-0300. doi: 10.1145/3618105. URL https://doi.org/10. 1145/3618105. Prado-Romero, M. A., Prenkaj, B., and Stilo, G. Gretel 2.0: Generation and evaluation of graph counterfactual expla- nations evolved. In Joint European Conference on Ma- chine Learning and Knowledge Discovery in Databases, pp. 363–367. Springer, 2024a. Prado-Romero, M. A., Prenkaj...

    doi:10.1145/3618105

  6. [12]

    URL https://doi.org/ 10.24963/ijcai.2021/214

    doi: 10.24963/ijcai.2021/214. URL https://doi.org/ 10.24963/ijcai.2021/214. Main Track. Srinivas, S., Subramanya, A., and Babu, R. Training sparse neural networks. In 2017 IEEE Conf. on CVPRW , pp. 455–462. IEEE,

    doi:10.24963/ijcai.2021/214 2021

  7. [13]

    J., O’brien, L

    Sutherland, J. J., O’brien, L. A., and Weaver, D. F. Spline- fitting with a genetic algorithm: A method for developing classification structure- activity relationships. Journal of chemical information and computer sciences , 43(6): 1906–1915,

    work page 1906

  8. [15]

    ISBN 9781450390965

    Association for Computing Machinery. ISBN 9781450390965. URL https://doi.org/10. 1145/3485447.3511948. Vinyals, O., Fortunato, M., and Jaitly, N. Pointer networks. Advances in neural information processing systems, 28,

    arXiv

Show all 27 references
  1. [16]

    Wu, X., Xiong, Y ., Zhang, Y ., Jiao, Y ., Shan, C., Sun, Y ., Zhu, Y ., and Yu, P. S. Clare: A semi-supervised commu- nity detection algorithm. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp. 2059–2069,

  2. [17]

    A short survey of recent advances in graph matching

    Yan, J., Yin, X.-C., Lin, W., Deng, C., Zha, H., and Yang, X. A short survey of recent advances in graph matching. In Proceedings of the 2016 ACM on international conference on multimedia retrieval, pp. 167–174,

    2016

  3. [19]

    We train GIST with Adam optimizer with learning rate 10−3 and a weight decay of 10−5

    We set α = 0.9 to encourage higher validity, which is beneficial for a helpful counterfactual. We train GIST with Adam optimizer with learning rate 10−3 and a weight decay of 10−5. • For CF2 (Tan et al., 2022), we configured: 20 epochs, batch size ratio of 0.2, learning rate (...

    2022

  4. [20]

    † depicts a real-world dataset and ‡ a synthetic one

    Summary of selected datasets and their characteristics. † depicts a real-world dataset and ‡ a synthetic one. Dataset Graphs Classes Avg. Nodes Avg. Edges Node Labels Edge Labels Node Attr. AIDS † 2000 2 15.69 16.20 ✓ ✓ ✓ (4) BAShapes ‡ 500 2 57.00 82.01 – – ✓(8) BBBP † 2039 2...

    2000

  5. [21]

    Following the generation done in (Lucic et al., 2022), there are 8 nodes on the base graph with 5 edges connecting them

    or a grid-shape (class 1). Following the generation done in (Lucic et al., 2022), there are 8 nodes on the base graph with 5 edges connecting them. Each base graph has 7 motives connected to it. • BBBP (Blood-Brain Barrier Penetration) (Martins et al.,

    2022

  6. [23]

    Each image is represented by a conditional Markov random field graph, as illustrated in Figure

    is a state-of-the-art dataset in semantic image processing originally introduced in (Winn et al., 2005). Each image is represented by a conditional Markov random field graph, as illustrated in Figure

    2005

  7. [25]

    The left-most RGB image is represented by a graph of superpixels (middle) with semantic labels b building, c car, v void, and ? unlabeled

    Taken from (Neumann et al., 2016). The left-most RGB image is represented by a graph of superpixels (middle) with semantic labels b building, c car, v void, and ? unlabeled. (right) Point clouds of household objects represented by labeled 4-nearest-neighbor graphs with part la...

    2016

  8. [26]

    This metric, akin to runtime, assesses the computational complexity of the explainer, especially in distributed systems

    quantifies the number of times the explainer queries the oracle to produce a counterfactual. This metric, akin to runtime, assesses the computational complexity of the explainer, especially in distributed systems. It avoids considering latency and throughput, which are externa...

    2022

  9. [27]

    with the hyperparameter search space introduced in the original paper does not reach more than 40% of accuracy in the test set. We ran experiments with these oracles for NCI1, however the fidelity of the explainers was negative, which suggests that the explainers are actually ...

    2023

  10. [28]

    Among all methods, GIST achieves the highest validity (0.87) and fidelity (0.17) w.r.t

    Notably, we re-ran the CLEAR – the original paper that uses IMDB-M – from scratch and obtained a validity score of 0.45, which significantly deviates from the 0.96 originally reported, raising questions about the reproducibility and robustness of the original evaluation. Among...

    2022

  11. [2003]

    Learning and evaluating graph neural net- work 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 net- work explanations based on counterfactual and factual reasoning. In Proceedings of the ACM Web Confer- ence 2022 , WWW ’22, pp. 1018–1027, New York, NY , USA,

    2022

  12. [2005]

    The dataset includes 600 enzymes from the BRENDA database (Schomburg et al., 2004), with 100 proteins sampled from each of the six Enzyme Commission (EC) top-level classes

    consists of protein graph models. The dataset includes 600 enzymes from the BRENDA database (Schomburg et al., 2004), with 100 proteins sampled from each of the six Enzyme Commission (EC) top-level classes. The original objective that the dataset was created for is to accurate...

    2004

  13. [2016]

    Q., Nguyen, T

    Nguyen, D. Q., Nguyen, T. D., and Phung, D. Universal graph transformer self-attention networks. In Companion Proceedings of the Web Conference 2022, pp. 193–196,

    2022

  14. [2018]

    A., Bethge, M., Hertzmann, A., and Shechtman, E

    Gatys, L. A., Bethge, M., Hertzmann, A., and Shechtman, E. Preserving color in neural artistic style transfer. arXiv preprint arXiv:1606.05897, 2016a. Gatys, L. A., Ecker, A. S., and Bethge, M. Image style trans- fer using convolutional neural networks. In Proceedings of the I...

    arXiv

  15. [2019]

    Kosan, M., Huang, Z., Medya, S., Ranu, S., and Singh, A

    URL http://arxiv.org/ abs/1905.02850. Kosan, M., Huang, Z., Medya, S., Ranu, S., and Singh, A. Gcfexplainer: Global counterfactual explainer for graph neural networks. ACM Transactions on Intelligent Systems and Technology,

    1905 arXiv

  16. [2021]

    A., Prenkaj, B., and Stilo, G

    Prado-Romero, M. A., Prenkaj, B., and Stilo, G. Devel- oping and evaluating graph counterfactual explanation with gretel. In Proceedings of the Sixteenth ACM Interna- tional Conference on Web Search and Data Mining, pp. 1180–1183, 2023a. Prado-Romero, M. A., Prenkaj, B., and S...

    2023

  17. [2022]

    and Bacciu, D

    Numeroso, D. and Bacciu, D. Meg: Generating molecular counterfactual explanations for deep graph networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pp. 1–8. IEEE,

    2021

  18. [2023]

    Regulation (eu) 2016/679 of the european parliament and of the council

    European Union. Regulation (eu) 2016/679 of the european parliament and of the council. Official Journal of the European Union,

    2016

  19. [2024]

    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

Pith tools

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