acceptodds
Withdrawn from ICLR 2027

Representations + Lie Algebra = Explanations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph Neural Networks (GNNs) are powerful yet remain `black box' models, and their use in high-stakes domains requires explanation methods that are both mathematically principled and semantically faithful. Existing factual explainers identify influential subgraphs via discrete combinatorial search, while counterfactual methods perturb the latent embedding without geometric constraints, frequently producing out-of-distribution (OOD) artifacts that are infeasible or structurally invalid. We address both limitations by framing GNN reasoning as symmetry transformation on a learned latent manifold. To this end, we introduce GLIDE: Graph LIe Decomposition Explainer. We model the GNN latent space as a smooth manifold acted upon by a continuous Lie group , and we learn a basis for the associated Lie algebra . Factual explanations correspond to the label-stabilizer orbit (group actions generated by the subalgebra that leave the predicted class invariant). Counterfactual explanations correspond to transverse flows generated by the complement , which traverse the decision boundary along the shortest algebraic path. We prove that this decomposition is reductive, guarantee that all generated trajectories remain within a bounded neighborhood of the data manifold, establish that orthogonal generator bases produce locally independent counterfactual directions, and show that winner-take-all sparsity recovers the unique minimal-norm counterfactual generator. Empirical evaluation on synthetic (BA-2Motifs) and biochemical (MUTAG, PROTEINS, NCI1, AIDS) benchmarks demonstrates that our framework achieves high-quality factual and counterfactual explanations with strict manifold proximity, outperforming all baseline methods.

Primary area:
interpretability and explainable AI
Keywords:
GNN Explainability, Factual and Counterfactual Explanation, Interpretability, Geometric Deep Learning, Lie Algebra, Representation Geometry, Symmetry Transformations & Group Actions

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_eDXvOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a Lie-algebraic framework for unifying factual and counterfactual GNN explanations. It learns low-rank generators to produce label-preserving and label-altering transformations in a frozen latent space, with a coefficient network controlling the transformation magnitude. The authors evaluate on five graph datasets, report ablation studies, and include assumption-check experiments.

Critical strengths

The paper presents a detailed pipeline combining low-rank latent transformations, coefficient prediction, and staged training for factual and counterfactual explanations.

Critical weaknesses

Definition 1 appears not to define a subgroup. Label preservation at a single point is not closed under composition, and at a strict-margin sample the set contains a neighborhood of the identity, so its tangent space becomes the entire Lie algebra. This leaves no nontrivial one-dimensional counterfactual complement, which directly affects the claimed decomposition and the completeness argument of Theorem 2.

I also note that the mechanism experiments mainly measure boundary-crossing and sparsity behavior, which reflect properties of the training protocol rather than the structural conditions assumed in the theorems. They are useful diagnostics, but they do not provide independent validation of the theoretical claims.

Overall, the ambition of this work is a strength. Connecting representation-space transformations with factual and counterfactual explanations through Lie-theoretic structure is an intellectually compelling direction. However, I would appreciate clarification on the foundational issues identified above.

Questions for authors

For a point with a strict classification margin, how can a nontrivial factual subalgebra and a one-dimensional counterfactual complement arise, and how does this affect the completeness claim in Theorem 2 and the minimal-generator claim in Theorem 4?

7 comments
AuthorsOpenReview ↗

Response to Reviewer eDXv part 1 of 2

(Response to reviewer eDXv part 1/2)

We thank the reviewer for their constructive feedback and precise questions. We appreciate that the reviewer found our pipeline, which combines low-rank latent transformations, coefficient prediction, and staged training for factual and counterfactual explanations, detailed, and that they considered the ambition of the work a strength, describing the connection between representation-space transformations and factual and counterfactual explanations through Lie-theoretic structure as an intellectually compelling and ambitious direction. We address each point below, and have marked all corresponding updates in blue in the revised manuscript.

W1.1 (Definition 1 appears not to define a subgroup)

The reviewer is correct that HzH_z is not a subgroup. We did in fact initially have a paragraph discussing that HzH_z is not formally a subgroup because it is not closed, etc., however it was commented out by mistake due to space constraints and as a result this confusion has persisted and affected the “soundness” evaluation of the paper. We apologize for this, and for the incorrect use of the term; in fact we do not need any closure guarantees. This is both correct and by design, and it does not affect the decomposition or any downstream results because none of these use or need subgroup properties. HzH_z is not intended to be a subgroup, nor does it need to be. Hz=g:y^(g⋅z)=y^(z)H_z = {g: \hat{y}(g \cdot z) = \hat{y}(z)} is the preimage of the class region under the orbit map g↦g⋅zg \mapsto g \cdot z. Membership of g1g_1 is tested at zz and we do not claim that it also applies at g2⋅zg_2 \cdot z. HzH_z can be understood as the set of label-preserving transformations of one sample (the latent representation zz of an input graph xx). This does not mean that it is a symmetry group of the classifier as a whole.

Moreover, we do not want HzH_z to be a subgroup because at a strict margin zz, HzH_z contains an open neighborhood UU of ee. A subgroup containing UU contains the identity component G0G_0, which is generated by UU. If HzH_z were then to be a subgroup, every g∈G0g \in G_0 would preserve the label y^(z)\hat{y}(z), making no counterfactual reachable. As such, HzH_z can only be a true subgroup when there is only one class (nothing to explain).

HzH_z also does not need to be a subgroup for any of our results. The closure is not necessary for our method, theoretical results, or interpretability claims. The method uses only the linear subspace hz=kerℓz\mathfrak{h}_z = ker \ell_z and the flows t↦exp⁡(tX)zt \mapsto \exp(tX)z, which are one-parameter subgroups of GG, not of HzH_z.

The codimension-one decomposition is proved independently. Lemma 1 gives dim⁡kerℓz=dim⁡g−1\dim ker \ell_z = \dim \mathfrak{g}-1 from Assumption 1 (ℓz≢0\ell_z \not\equiv 0) alone. It does not use subgroups, faithfulness, or surjectivity assumptions. Equation 4 decomposes g=kerℓz⊕mz\mathfrak{g} = ker \ell_z \oplus \mathfrak{m_{\text{z}}} from it, and training targets kerℓzker \ell_z directly with Ltan⁡\mathcal L_{\tan}.

For Theorems 1 and 2, the local factorization g=exp⁡(Xfact)exp⁡(Xcf)g=\exp(X_{\rm fact})\exp(X_{\rm cf}) holds for any complementary pair (inverse function theorem), so it needs no subgroup. Reductivity is a property of the learned span that is enforced by Lclosure\mathcal L_{\rm closure} and Ladj\mathcal L_{\rm adj}. ker⁡ℓz\ker\ell_z does not need bracket closure in general. The leaves of Theorem 2 are the level sets of yy that foliate Zy\mathcal{Z}_y (without any group structure). We will restate completeness only for exp⁡(hz)z\exp(\mathfrak h_z)z rather than for the entire set.

Theorems 3 – 5 and Propositions 2 – 3 do not use HzH_z. Theorem 2 has been restated by level sets of yy and this does not need group structure.

W1.2 (Where the nontrivial factual subalgebra and 1-D complement come from)

They come from the logit level set Lz=g:y(g⋅z)=y(z)⊂HzL_z = {g:y(g \cdot z) = y(z)} \subset H_z, not from TeHzT_eH_z.

Let F(g) = y(g \cdot z). Since dFe=ℓz≠0dF_e=\ell_z\neq0, the regular value theorem makes LzL_z a codimension-one submanifold near ee, with TeLz=ker⁡ℓz=:hzT_eL_z=\ker\ell_z=:\mathfrak{h}_z. So mz=hz⊥\mathfrak{m}_z=\mathfrak{h}_z^\perp is one-dimensional. Elements of hz\mathfrak{h}_z preserve the logit to first order, y(exp(tX)z−y(z)=O(t2)y(exp(tX)z - y(z) = O(t^2), and hence preserve the label for small t. mz\mathfrak m_z is the unique first-order logit-changing direction. The revision will define hz\mathfrak h_z this way and Proposition 1's faithfulness and local-transitivity hypotheses are replaced by ℓz≢0\ell_z\not\equiv0.

AuthorsOpenReview ↗

Response to Reviewer eDXv part 2 of 2

(Response to reviewer eDXv part 2/2)

W1.3 (Theorems 2 and 4)

We will restate this as completeness under the orbit, not over every point. The factual flows reach every point of the local orbit exp⁡(hz)z\exp(\mathfrak{h}_z)z, not every nearby point of Zy\mathcal{Z}_y. When ρz\rho_z is injective on hz\mathfrak{h}_z, this orbit is an immersed submanifold of dimension dim⁡hz\dim\mathfrak{h}_z. It covers the level set of y only if dim⁡g≥D\dim\mathfrak g\ge D, which our K-generator span does not satisfy. The foliation itself is by level sets of y and needs no group structure.

Theorem 4. Its form is unaffected. It uses only dim⁡mz=1\dim\mathfrak m_z=1, which now follows from ℓz≢0\ell_z\not\equiv0. It remains a first-order statement restricted to one-hot coefficients (Section 8).

W2 (Mechanism experiments)

Our theorems are conditional statements. Given their hypotheses, the conclusions follow by proof and need no empirical validation. The hypotheses are first-order tangency of the factual generators, bracket closure, the reductive condition, and concentrated (one-hot) coefficients. What can fail in practice is whether a trained model satisfies those hypotheses.

GLIDE is designed so that it has the following properties, and as per Table 4, each loss term enforces one condition:

  • Ltan\mathcal{L_{\text{tan}}} enforces tangency, Af,kz∈ker∇zy⊤A_{f, k}z \in ker \nabla_z y^{\top}.

  • Lclosure\mathcal{L_{\text{closure}}} enforces bracket closure of the factual span.

  • Ladj\mathcal{L_{\text{adj}}} enforces the reductive condition [h,m]⊆m[\mathfrak{h},\mathfrak{m}]\subseteq\mathfrak{m}.

  • Lentropy\mathcal{L_{\text{entropy}}} encourages the concentrated coefficients assumed by Thm. 4.

  • Lmargin\mathcal{L_{\text{margin}}} enforces the label flip of Def. 3.

The mechanism experiments verify that this construction succeeds. The hypotheses check that the learned generators meet the theorems' hypotheses on held-out test graphs, where they were not optimized. They are not independent evidence for the theorems themselves, but they show that we are indeed operating under this setting.

We hope this review answers all of the reviewer's concerns, and we are happy to address any remaining questions during the discussion period.

I appreciate the corrections, but they have not resolved my concerns. I am lowering my rating to clear rejection.

My main concern is that the authors have weakened the theoretical results without reconsidering the claims built on them. The subgroup has become a first-order kernel, completeness has been restricted, and manifold preservation has become a displacement bound. These changes affect the substance of the contribution. The revised paper still needs to show why the remaining results support its claims about explanation quality.

There are also unresolved contradictions. Definition 1 says that HzH_z is not a subgroup, yet the next paragraph identifies its connected component's tangent space with the factual Lie algebra. Corollary 1 still uses G/Hz0G/H_z^0 as a homogeneous space (pp. 3-5). At a strict-margin point, HzH_z contains an open neighborhood of the identity, so its tangent space is the full algebra. Separately, the first-order tangency in Proposition 1 does not guarantee margin preservation along finite exponential flows. I recognize that the new displacement bounds are tighter. This does not answer the manifold concern. They bound distance from the original embedding, and even an arbitrarily small displacement can leave the data manifold. Making this bound tighter cannot establish manifold membership or the claimed control of OOD drift. Without an additional argument linking displacement to data validity, those claims need to be withdrawn and the narrower contribution evaluated on its own merits.

The response about mechanism experiments also misses my concern. I am not asking for empirical proof of a theorem. Checking assumptions on held-out data is useful, but training penalties do not guarantee that the assumptions hold exactly. The paper needs to account for the remaining errors. Likewise, observing sparsity after hard top-1 masking does not show that the selected generator minimizes displacement. Removing a loss term can show that it helps, but does not by itself explain whether the gain comes from the proposed geometry or ordinary regularization.

Theorem 4 is another example. It assumes geometric convergence and then derives convergence of the displacement quantities. The decreasing envelopes do not mean that the actual displacement or computed bounds decrease during training. Lemma 2 gives sufficient conditions involving local strong convexity, fixed generator selection, and full-batch gradient descent. These have not been established for the actual mini-batch Adam training on low-rank factors with a straight-through Gumbel mask. The result therefore does not explain why the implemented training should behave as claimed.

The response still leaves unclear how a fair comparison with the baselines was ensured and whether the results were rechecked after the evaluation code was corrected. The metrics themselves also need care. Zero generators give Fid+ = 0 and Fid- equal to the original prediction confidence, so a confident classifier can score very well without producing a meaningful explanation. The motif results are useful, but do not settle this concern about the broader comparisons.

These issues require more than another clarification of the definitions. The authors need to rebuild the paper around the claims they can support and design experiments that test those claims.

Reviewer_JSx2OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes GLIDE, Graph Lie Decomposition Explainer. GLIDE is a post-hoc explainer. The explained model is a GCN or GIN encoder with a single linear classifier head, pre-trained and then frozen.

GLIDE learns KK matrices on the encoder's latent space, in low-rank form Ak=UkVk⊤A_k = U_kV_k^\top. A coefficient network predicts per-graph weights tt, and the transformation is applied by the matrix exponential, z′=exp⁡(∑ktkAk),zz' = \exp(\sum_k t_kA_k),z. A factual explanation is the orbit of the matrices that preserve the predicted label. A counterfactual is a flow along the complement, across the decision boundary. Six theorems and three propositions support the decomposition, a proximity bound and a sparsity result.

Datasets: BA-2Motifs, MUTAG, PROTEINS, NCI1, AIDS. Factual baselines, counterfactual baselines and two symmetry-discovery methods are compared. Code is supplied in the supplement.

Critical strengths
  1. The label-preserving / label-altering split is new. I found no prior work that partitions a learned Lie algebra by its effect on a classifier's decision. Treating factual explanation (label-preserving motion) and counterfactual explanation (boundary-crossing motion) as complementary halves of one structure is a good framing.
Critical weaknesses
  1. The theory part is very hard to assess and verify.
  • The main text is not self contained, multiple algebra notions are undefined.

  • At a fixed zz the whole content of Section 3 is one linear functional on a space of matrices. With ℓz(A)=∇zy(z)⊤(Az)\ell_z(A)=\nabla_zy(z)^\top(Az), Proposition 1(ii)–(iii) say that ker⁡ℓz\ker\ell_z has codimension one. That is rank–nullity.

  • The geometry is a hyperplane. The released classifier is nn.Linear(latent_dim, num_classes), so y(z)=wc⊤z+bcy(z)=w_c^\top z+b_c and ∇zy(z)=wc\nabla_zy(z)=w_c — a constant vector, fixed after pre-training. The decision boundary is an affine hyperplane in R64\mathbb R^{64} and its level sets are parallel translates. Smoothness, local structure and manifold constraint are being asserted about an affine map.

  • Theorem 1 seems to restate a definition. Is there any novelty beyond what exists in a differential geometry textbook?

  • Theorem 3's bound is vacuous. On the paper's own numbers (p33, ∣B∣F=8.845|B|_F=8.845, mean ∣z∣=3.282|z|=3.282) it evaluates to 3.282(e8.845−1)≈2.3×1043.282(e^{8.845}-1)\approx 2.3\times10^4, about 6,900 times the radius of the latent data.

  1. The released code has multiple issues.
  • The factual metrics apply the counterfactual generators. Both branches of SymmetryMetrics.calculate_explanation_fidelity call model.cf_gens; fact_gens appears nowhere in that file. evaluate_full calls it with mode="factual" to produce the Fid+ fact / Fid- fact values that runner.py scrapes. If that is the path behind Table 1, the factual generators trained over 300 epochs in Phase 1 are never evaluated by any reported factual metric.

  • The method and the baselines are scored under different definitions of a counterfactual. For GLIDE the transformation is conditioned on the target class, t = coef_net(cat([z, one_hot(target_y)])). For the baselines it is a constant, t = model.coef.expand(B,-1), independent of the input and of the target. Both are then scored on P(targety)P({target_y}).

  • No baseline is reproducible. The archive contains no GNNExplainer, PGExplainer, RCExplainer, CLEAR, D4Explainer or GCFExplainer. LieGAN and Augerino⁺ appear only as helpers consuming objects supplied from outside. No column of Tables 1 or 2 other than GLIDE's can be checked.

Questions for authors
  1. In Section 3.1, can you explain why is hz\mathfrak h_z closed under the Lie bracket?

  2. Does Theorem 1 establishes anything that Eq. (5) does not already assume?

  3. It appears that the method and the baselines are evaluated under different definitions of a counterfactual. In SymmetryMetrics.calculate_explanation_fidelity, GLIDE's transformation is conditioned on the target class while the baselines receive a constant vector, and both are then scored on the probability of that same target class. Could you check?

  4. Which generator head produced the factual numbers, and which objective was run? Both branches of calculate_explanation_fidelity call model.cf_gens, so the reported factual fidelity appears to apply the counterfactual generators.

3 comments
AuthorsOpenReview ↗

Response to Reviewer JSx2 part 1 of 2

(Response to reviewer JSx2 part 1/2)

We thank the reviewer for their detailed reading of both the paper and the code. We appreciate that the reviewer found the label-preserving / label-altering split new, noting that they found no prior work that partitions a learned Lie algebra by its effect on a classifier's decision, and that they found treating factual explanation (label-preserving motion) and counterfactual explanation (boundary-crossing motion) as complementary halves of one structure a good framing. We address each point below, and have marked all corresponding updates in blue in the revised manuscript.

W1.1

We have included a list of notations used in the paper in Appendix C and a background on Lie Algebra in Appendix D as clarifying and supplementary material that is needed to understand the concepts in the paper. If there is another concept the reviewer believes is necessary to understand the paper (but is not sufficiently discussed) we would be happy to add supplementary information on this.

W1.2

The reviewer is correct that the co-dimension one claim is simply rank-nullity and only needs ℓz≢0\ell_z \not\equiv 0. Proposition 1 has been updated (as per comments from Reviewer ABNk) to: hz:=kerℓz\mathfrak h_z := ker \ell_z has codimension one iff ℓz≢0\ell_z\not\equiv0. The contribution lies in the application of this lemma, that a supervised split of a learned generator span into first-order label-preserving and label-changing parts. As you note in the strengths section of your review, this label-preserving / label-altering split is new, and not studied this way by prior work.

W1.3

This point is correct. hϕh_{\phi} is linear, making ∇zy=wc\nabla_z y = w_c constant, and the boundary and level sets then parallel affine hyperplanes in R64\mathbb R^{64}. In fact, this simplifies Proposition 1(iv) and Theorem 4 (along with the corresponding Section 8 limitation where the label-flip condition becomes exact and the remaining first-order error comes only from the exponential map), making them more concise/ strengthening them. The group action still depends on zz since ℓz(A)=wc⊤Az\ell_z(A) = w_c^\top Az, hz\mathfrak h_z varies with zz, and the flows exp(tA)zexp(tA)z are curved.

W1.4

The reviewer is correct here. The local factorization holds for any complementary pair by the inverse function theorem, and Eq. 5 assumes the reduction. We agree it would be more fitting to reduce Theorem 1 to a proposition and claim no novelty for it.

Two related points are specific to our setting. The choice of the complement as hz=ker⁡ℓz\mathfrak h_z=\ker\ell_z is the space of first-order label-preserving directions, and dim⁡mz=1\dim\mathfrak m_z=1 (Lemma 1). The two factors are therefore a label-preserving move and a label-changing move, which is what motivates the two phase design. Secondly, the reductivity is not automatic for learned generators, and we enforce it with Ladj\mathcal L_{\rm adj}.

W1.5

The computation here is correct. The bound ∣z∣(eM∣t∣1−1)|z|(e^{M|t|_1}-1) holds for arbitrary matrices, so it is vacuous at our generator norms. We have updated Theorem 3 from “Manifold Proximity Guarantee" to “Latent Displacement Bound,” and we clarify that the proximity results reported in Table 2 are measured directly and do not depend on the bound. We strengthened the theorem with two bounds that use the structure of the learned generators instead of their norms. The revised Theorem 3 is updated in the manuscript pdf.

We also measured these quantities and reported them in Appendix N.3.

Measured over 10 runs (median run; per-run median over test graphs; full table with interquartile ranges in App. N.3):

BA-2MotifsMUTAGPROTEINSNCI1AIDS
Observed ∥z′−z∥2\|z'-z\|_25.803.772.892.341.77
First-order step ∥Bz∥2\|Bz\|_280.526.111.210.19.60
(i) worst case2.2×1092.2{\times}10^{9}3.7×1073.7{\times}10^{7}1.5×10101.5{\times}10^{10}3.0×10183.0{\times}10^{18}8.4×1048.4{\times}10^{4}
(iii) low-rank bound222821323629941.5
Observed / tightest bound0.0060.0320.0180.0100.055

The refined bounds are 3–16 orders of magnitude tighter than the worst case, which is vacuous on every dataset: by factors of about 2×1062\times 10^6 (BA-2Motifs), 2×1052\times 10^5 (MUTAG), 6×1076\times 10^7 (PROTEINS), 101610^{16} (NCI1) and 2×1032\times 10^3 (AIDS). They are within a factor of about 18–170 of the observed displacement: 170 (BA-2Motifs), 31 (MUTAG), 56 (PROTEINS), 100 (NCI1) and 18 (AIDS). They capture the order of growth but are not numerically tight. The displacement itself is given exactly by (iii), z′−z=U~φ(S)V~⊤zz'-z=\tilde U\varphi(S)\tilde V^\top z, and that exact quantity is what Table 2 reports.

AuthorsOpenReview ↗

Response to Reviewer JSx2 part 2 of 2

(Response to reviewer JSx2 part 2/2)

W2.1–3

We sincerely apologize for this and have realized that we have submitted a zip file from an earlier version of the code without the most current local branches merged. We have updated the zip file with the current code. We would like to clarify that the factual generators are frozen after phase 1 and are evaluated by the factual fidelity metrics, and that we score all methods with the target-conditioned coefficient network (trained on each baseline’s frozen generators). The other baseline methods have publicly available code that we ran out-of-the-box, and all are referenced within our paper.

Q1

The reviewer is correct that HzH_z is not a subgroup. We did in fact initially have a paragraph discussing that HzH_z is not formally a subgroup because it is not closed, etc., however it was commented out by mistake due to space constraints and as a result this confusion has persisted and affected the “soundness” evaluation of the paper. We apologize for this, and for the incorrect use of the term; in fact we do not need any closure guarantees. This is both correct and by design, and it does not affect the decomposition or any downstream results because none of these use or need subgroup properties. HzH_z is not intended to be a subgroup, nor does it need to be. Hz=g:y^(g⋅z)=y^(z)H_z = {g: \hat{y}(g \cdot z) = \hat{y}(z)} is the preimage of the class region under the orbit map g↦g⋅zg \mapsto g \cdot z. Membership of g1g_1 is tested at zz and we do not claim that it also applies at g2⋅zg_2 \cdot z. HzH_z can be understood as the set of label-preserving transformations of one sample (the latent representation zz of an input graph xx). This does not mean that it is a symmetry group of the classifier as a whole.

Moreover, we do not want HzH_z to be a subgroup because at a strict margin zz, HzH_z contains an open neighborhood UU of ee. A subgroup containing UU contains the identity component G0G_0, which is generated by UU. If HzH_z were then to be a subgroup, every g∈G0g \in G_0 would preserve the label y^(z)\hat{y}(z), making no counterfactual reachable. As such, HzH_z can only be a true subgroup when there is only one class (nothing to explain).

HzH_z also does not need to be a subgroup for any of our results. The closure is not necessary for our method, theoretical results, or interpretability claims. The method uses only the linear subspace hz=kerℓz\mathfrak h_z = ker \ell_z and the flows t↦exp⁡(tX)zt \mapsto \exp(tX)z, which are one-parameter subgroups of GG, not of HzH_z.

The codimension-one decomposition is proved independently. Lemma 1 gives dim⁡kerℓz=dim⁡g−1\dim ker \ell_z = \dim \mathfrak{g}-1 from Assumption 1 (ℓz≢0\ell_z \not\equiv 0) alone. It does not use subgroups, faithfulness, or surjectivity assumptions. Equation 4 decomposes g=kerℓz⊕mz\mathfrak{g} = ker \ell_z \oplus \mathfrak m_z from it, and training targets kerℓzker \ell_z directly with Ltan⁡\mathcal L_{\tan}.

For Theorems 1 and 2, the local factorization g=exp⁡(Xfact)exp⁡(Xcf)g=\exp(X_{\rm fact})\exp(X_{\rm cf}) holds for any complementary pair (inverse function theorem), so it needs no subgroup. Reductivity is a property of the learned span that is enforced by Lclosure\mathcal L_{\rm closure} and Ladj\mathcal L_{\rm adj}. ker⁡ℓz\ker\ell_z does not need bracket closure in general. The leaves of Theorem 2 are the level sets of yy that foliate Zy\mathcal{Z}_y (without any group structure). We will restate completeness only for exp⁡(hz)z\exp(\mathfrak h_z)z rather than for the entire set.

Theorems 3 – 5 and Propositions 2 – 3 do not use HzH_z. Theorem 2 has been restated by level sets of yy and this does not need group structure.

We hope this review answers all of the reviewer's concerns, and we are happy to address any remaining questions during the discussion period.

I thank the authors for the careful response. Every technical point in my review has been integrated: that the codimension-one claim is rank-nullity, that the classifier is linear so the level sets are parallel affine hyperplanes, that Theorem 1 should be demoted with no novelty claimed, that Theorem 3's bound is vacuous at the reported generator norms, and that HzH_z is not a subgroup.

I am maintaining my score, for two reasons.

First, three of my weaknesses are not addressed:

  • W1.1, self-containment. The response refers to a notation list in Appendix C and Lie-algebra background in Appendix D. My concern was that the main text is not self-contained, and moving further material into appendices does not address it.

  • W2.1 and Q4, which code produced the results. The response states that the archive was an earlier version, and that the factual generators are frozen and evaluated by the factual fidelity metrics. It does not state which version produced Tables 1 and 2. If the experiments were run with the submitted code, the reported factual fidelity applied the counterfactual generators.

  • W2.3, baseline reproducibility. That the baselines have public implementations does not make the reported columns checkable. The archive contains no configuration, hyperparameters or adaptation for any of them, and running a factual explainer out of the box against a frozen target model and these splits is not a neutral operation.

Second, the changes are substantial rather than clarifying. Proposition 1 is restated, Theorem 1 becomes a proposition with no novelty claimed, Theorem 2 is restated in terms of level sets, Theorem 3 is renamed and replaced with two new bounds, Theorem 4 and Proposition 1(iv) are simplified, the Section 8 limitation is rewritten, three appendices are added, and the submitted code was an earlier version. A discussion period is for clarifying a submitted paper, not for assessing a replacement of it, and new theorems, new bounds and new code cannot be verified within one.

I continue to think the framing deserves a correct paper, and I would encourage resubmission of the corrected version, with the theory stated at the strength it has and the evaluation run on code that matches it.

Reviewer_ABNkOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes GLIDE, a GNN explanation framework that unifies factual and counterfactual explanations through Lie algebraic transformations in the latent representation space. The authors model the GNN latent space as a manifold equipped with a Lie group action and decompose the associated Lie algebra into label-preserving and label-changing components. Factual explanations are characterized through label-stabilizing transformations, while counterfactual explanations are obtained by traversing complementary directions using the matrix exponential. The paper further develops several theoretical results concerning reductive decomposition, manifold proximity, generator orthogonality, minimal counterfactual directions, and two-phase training dynamics. Experiments on five graph classification benchmarks, including BA-2Motifs, MUTAG, PROTEINS, NCI1, and AIDS, report favorable factual and counterfactual explanation performance compared with existing methods. Additional experiments investigate the proposed theoretical properties and the contributions of individual loss components.

Critical strengths
  • S1. An interesting approach to connecting factual and counterfactual GNN explanations through Lie theory. The paper presents a different perspective on GNN interpretability by modeling factual and counterfactual explanations as transformations within a shared latent space. Using Lie algebra generators to characterize label-preserving and label-changing transformations is an interesting idea that may offer a useful way to study both types of explanations. The connection drawn between explanation directions and the geometry of classifier decision boundaries is particularly worth exploring.

  • S2. A fairly comprehensive experimental evaluation with supporting theoretical analysis. The authors evaluate GLIDE on five graph classification datasets, covering both factual and counterfactual explanations. In addition to the main comparisons, the paper includes ablation studies, visualizations, and further experiments examining its theoretical claims. The appendices also provide detailed mathematical derivations and experimental settings, which are helpful for understanding the proposed method and assessing the validity of its claims.

Critical weaknesses
  • W1. Fundamental mathematical problems cast doubt on the proposed Lie-theoretic formulation and its theoretical guarantees. My main concern is that several key mathematical arguments are either insufficiently justified or potentially incorrect.

    • First, the label-stabilizer introduced in Definition 1 is not necessarily a subgroup. The authors define it as Hz={g∈G:y^(g⋅z)=y^(z)}H_z=\{g\in G:\hat y(g\cdot z)=\hat y(z)\}. However, transformations that preserve the predicted label at a given point are not generally closed under group composition. This calls into question whether the corresponding label-stabilizer Lie algebra is well-defined and whether Proposition 1 holds.
    • Second, Proposition 1 and its proof seem to treat label preservation as equivalent to logit invariance, although these are different properties. A transformation can preserve the predicted class while changing its logit value, including at first order. Yet Appendix H.2 relies on this assumption to derive hz⊆ker⁡ℓzh_z\subseteq\ker\ell_z. Another issue concerns the evaluation map ρz\rho_z: the proof uses its injectivity to argue that an arbitrary vector in the representation space can be obtained from some generator. This would require surjectivity, which injectivity alone does not guarantee. Consequently, the claimed equivalence between the factual subalgebra and the kernel of the label differential, along with the resulting codimension-one decomposition, is not convincingly established.
    • Third, Theorem 3 does not appear to justify the claimed manifold proximity guarantee. Eq. (8) only bounds the Euclidean distance ∥z′−z∥2\|z'-z\|_2 between the original and transformed representations. Being close to the original embedding, however, does not guarantee that the transformed point remains on the data manifold or corresponds to a realizable graph. Similarly, although the matrix exponential produces an invertible transformation, this property alone does not ensure that the transformation preserves the manifold of valid graph representations.
  • W2. The experimental results do not provide sufficient evidence for the claimed explanation quality or manifold preservation. Although the reported performance appears promising, several evaluation choices make it difficult to assess the strength of the conclusions.

    • First, GLIDE produces counterfactual explanations in the latent space rather than generating actual counterfactual graphs. As discussed in Section 8 and Appendix M, the method does not reconstruct valid graph structures from the transformed representations. This differs fundamentally from the counterfactual baselines, which produce modified graphs through explicit editing or generation. As a result, it is unclear whether the methods are being compared under equivalent requirements.
    • Second, the definition of CF Proximity is inconsistent across the paper. Section 6.2 describes proximity as Euclidean distance in the latent space for all methods, whereas Appendix M.1 introduces normalized graph edit distance and notes that GLIDE is evaluated using embedding-space distance instead. Since these metrics capture different notions of proximity, their direct comparison may not be meaningful.
    • Third, the factual explanation metrics provide limited evidence of faithfulness. A high label-preservation score indicates that the classifier's prediction remains stable under a transformation, but it does not necessarily show that the explanation identifies the graph structures responsible for that prediction. Therefore, the reported results alone are insufficient to establish that GLIDE produces faithful factual explanations.
Questions for authors
  • Q1. Under what additional assumptions does the set Hz={g∈G:y^(g⋅z)=y^(z)}H_z=\{g\in G:\hat y(g\cdot z)=\hat y(z)\} form a Lie subgroup? As currently defined, it is not necessarily closed under composition. Could the authors explain how Proposition 1 and Theorems 1–2 hold without this property? Also, in the proof of Proposition 1, what justifies treating preservation of the predicted label as local logit invariance, and how does injectivity of the evaluation map imply the required surjectivity?

  • Q2. Theorem 3 provides an upper bound on ∥z′−z∥2\|z'-z\|_2, but this alone does not ensure that z′z' lies on the manifold of valid graph embeddings. Could the authors explain why the learned matrix-exponential transformations should preserve this manifold, preferably with a formal justification? If such a guarantee cannot be established, would it be more appropriate to describe Theorem 3 as a bound on latent-space displacement rather than a guarantee of generating in-distribution counterfactuals?

  • Q3. Could the authors clarify the discrepancy in the definition of CF Proximity between Section 6.2 and Appendix M.1? Are all methods evaluated using the same latent-space distance, or do the graph-generating baselines use graph edit distance while GLIDE uses embedding-space distance? If different metrics are used, how can their proximity scores be directly compared?

2 comments
AuthorsOpenReview ↗

Response to Reviewer ABNk part 1 of 2

(Response to reviewer ABNk part 1/2)

We thank the reviewer for their careful reading and constructive feedback. We appreciate that the reviewer highlighted the strength of our method and found our approach of connecting factual and counterfactual GNN explanations through Lie theory an interesting perspective, in particular using Lie algebra generators to characterize label-preserving and label-changing transformations, and that they found the connection between explanation directions and the geometry of the classifier's decision boundary particularly worth exploring. We also appreciate that they found the experimental evaluation comprehensive, covering five datasets in both factual and counterfactual settings with ablations, visualizations, and experiments examining the theoretical claims, and that they found the appendix derivations and experimental details helpful for assessing the validity of our claims. We address each point below, and have marked all corresponding updates in blue in the revised manuscript.

W1.1/ Q1 (HzH_z is not a subgroup).

The reviewer is correct that HzH_z is not a subgroup. We did in fact initially have a paragraph discussing that HzH_z is not formally a subgroup because it is not closed, etc., however it was commented out by mistake due to space constraints and as a result this confusion has persisted and affected the “soundness” evaluation of the paper. We apologize for this, and for the incorrect use of the term; in fact we do not need any closure guarantees. This is both correct and by design, and it does not affect the decomposition or any downstream results because none of these use or need subgroup properties. HzH_z is not intended to be a subgroup, nor does it need to be. Hz=g:y^(g⋅z)=y^(z)H_z = {g: \hat{y}(g \cdot z) = \hat{y}(z)} is the preimage of the class region under the orbit map g↦g⋅zg \mapsto g \cdot z. Membership of g1g_1 is tested at zz and we do not claim that it also applies at g2⋅zg_2 \cdot z. HzH_z can be understood as the set of label-preserving transformations of one sample (the latent representation zz of an input graph xx). This does not mean that it is a symmetry group of the classifier as a whole.

Moreover, we do not want HzH_z to be a subgroup because at a strict margin zz, HzH_z contains an open neighborhood UU of ee. A subgroup containing UU contains the identity component G0G_0, which is generated by UU. If HzH_z were then to be a subgroup, every g∈G0g \in G_0 would preserve the label y^(z)\hat{y}(z), making no counterfactual reachable. As such, HzH_z can only be a true subgroup when there is only one class (nothing to explain).

HzH_z also does not need to be a subgroup for any of our results. The closure is not necessary for our method, theoretical results, or interpretability claims. The method uses only the linear subspace hz=kerℓz\mathfrak{h}_z = ker \ell_z and the flows t↦exp⁡(tX)zt \mapsto \exp(tX)z, which are one-parameter subgroups of GG, not of HzH_z.

The codimension-one decomposition is proved independently. Lemma 1 gives dim⁡kerℓz=dim⁡g−1\dim ker \ell_z = \dim \mathfrak{g}-1 from Assumption 1 (ℓz≢0\ell_z \not\equiv 0) alone. It does not use subgroups, faithfulness, or surjectivity assumptions. Equation 4 decomposes g=kerℓz⊕mz\mathfrak{g} = ker \ell_z \oplus \mathfrak{m_{\text{z}}} from it, and training targets kerℓzker \ell_z directly with Ltan⁡\mathcal L_{\tan}.

We would like to address the remaining related concerns the reviewer raised that do hold and need revisions. We do need to revise our use of injectivity instead of surjectivity in Proposition 1, as the reviewer mentioned. The original text relating kerℓzker \ell_z to TeHz0T_eH_z^0 is incorrect because HzH_z is open and TeHz=gT_eH_z = \mathfrak{g}. We have revised this to define hz\mathfrak{h_{\text{z}}} as hz:=kerℓz\mathfrak{h_{\text{z}}} := ker \ell_z. Then geometrically, hz=TeLz\mathfrak{h_{\text{z}}} = T_eL_z for the logit level set Lz=g:y(g⋅z)=y(z)⊂HzL_z = {g:y(g \cdot z) = y(z)} \subset H_z, and this is a co-dimension one submanifold.

For Theorems 1 and 2, the local factorization g=exp⁡(Xfact)exp⁡(Xcf)g=\exp(X_{\rm fact})\exp(X_{\rm cf}) holds for any complementary pair (inverse function theorem), so it needs no subgroup. Reductivity is a property of the learned span that is enforced by Lclosure\mathcal L_{\rm closure} and Ladj\mathcal L_{\rm adj}. ker⁡ℓz\ker\ell_z does not need bracket closure in general. The leaves of Theorem 2 are the level sets of yy that foliate Zy\mathcal{Z}_y (without any group structure). We will restate completeness only for exp⁡(hz)z\exp(\mathfrak h_z)z rather than for the entire set.

Theorems 3 – 5 and Propositions 2 – 3 do not use HzH_z. Theorem 2 has been restated by level sets of yy and this does not need group structure.

AuthorsOpenReview ↗

Response to Reviewer ABNk part 2 of 2

(Response to reviewer ABNk part 2/2)

W1.2/ Q1 (label/ logits and injectivity/ surjectivity)

Label preservation comes as a consequence. We do not claim that label preservation implies logit invariance, only the other direction.

As pointed out by Reviewer JSx2, the co-dimension one claim is simply rank-nullity and only needs ℓz≢0\ell_z \not\equiv 0. The reviewer is correct to point out that injectivity does not give the needed vector, which needs surjectivity. The required condition is ℓz≢0\ell_z \not\equiv 0, which is exactly Assumption 1 (App. H.3), under which, Lemma 1 gives dim⁡kerℓz=dim⁡g−1\dim ker \ell_z = \dim \mathfrak{g} - 1 by rank–nullity. The revised version will delete the injectivity step and state Proposition 1 under Assumption 1 alone.

W1.3/ Q2 (Theorem 3)

The Euclidean embedding space distance does in fact only bound the latent displacement and we agree that this is not a direct measure of the counterfactual representations being on the data manifold. We will rename this as the “Latent Displacement Bound” such that we are strictly correct in terms of this terminology. We used this measure as a proxy for the measure of OOD quantification. Prior methods do not have a similar metric which we could baseline against, so we adapted this metric as an approximation under the assumption that if an embedding is closer in latent space, that it is more apt to be a valid counterfactual. The previous methods have much larger Euclidean distances, suggesting more off-manifold/ displaced behavior. Also under the assumption that input space graph similarity is reflected (at least to some degree) by the latent space proximity, this metric shows that GLIDE produces the minimal counterfactual edits, which is one of the desired properties of a true counterfactual explanation.

W2.1 (latent versus graph counterfactuals)

The reviewer is correct that GLIDE produces latent counterfactuals along with node saliencies, but does not produce decoded input-space graphs. We have addressed this as a limitation in Section 8 and Appendix M as the reviewer notes, and we acknowledge this as a direction of future work. We do, however, believe that the metrics we have provided (counterfactual accuracy, and Euclidean latent distance) give substantial evidence to support that these are valid, close counterfactuals.

W2.2/ Q3 (CF proximity)

We would like to confirm that Table 2 uses the latent distance for every method. This is ∣fθ(G′)−z∣2|f_\theta(G')-z|_2 for baselines, and ∣z′−z∣2|z'-z|_2 for GLIDE. We had left the edit distance text in Appendix M.1 from an earlier version and missed this as an oversight on our part. This incorrect text has been removed and we apologize for the confusion it may have caused.

W2.3 (factual faithfulness)

The latent label preservation shows that the generators are sufficient and serve as full explanations of the class. This is a standard metric for explanation quality. We have additionally provided visualizations of graph-space explanatory subgraphs in Figure 3, showing that GLIDE finds the exact explanatory subgraphs on examples from BA-2Motifs which does have ground truth labels. Since there was not room to put all of these, we have stated in Appendix G that the top-5 nodes selected by the dominant factual generator overlap with the ground-truth motif exactly on 100% (for cycle) and 92% (for house) of test samples. All explanations that do not have full motif recovery identify ≥ 3 motif nodes. This shows that our explanations successfully find the ground truth motifs.

We hope this review answers all of the reviewer's concerns. We are happy to address any remaining questions during the discussion period.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General comment to all reviewers and the AC

General comment to all reviewers and the AC

We thank all three reviewers for their careful and constructive reviews. Reviewer ABNk found the connection between factual and counterfactual GNN explanations through Lie theory an interesting perspective, highlighted the link between explanation directions and decision-boundary geometry as worth exploring, and found the experimental evaluation comprehensive, with helpful appendix derivations. Reviewer JSx2 found the label-preserving / label-altering split new, noting that no prior work partitions a learned Lie algebra by its effect on a classifier's decision, and found treating factual and counterfactual explanation as complementary halves of one structure a good framing. Reviewer eDXv found the pipeline detailed and the ambition of the work a strength, describing the direction as intellectually compelling.

All three reviewers raised the same central point: the label-preserving set HzH_z in Definition 1 is not a subgroup. We agree, and have corrected the paper. HzH_z is the set of label-preserving transformations of one sample, not a symmetry group of the classifier. It should not be a subgroup as well because at a strict margin it contains a neighbourhood of the identity, so a subgroup would contain the whole identity component, and no counterfactual could exist. No result depends on it being a subgroup. The method uses the factual subspace hz=ker⁡ℓz\mathfrak h_z=\ker\ell_z and one-parameter flows of GG. The codimension-one decomposition follows from Lemma 1 under the single assumption ℓz≢0\ell_z\not\equiv0, and geometrically hz\mathfrak h_z is the tangent space of the logit level set. In the revised PDF, changes are shown in blue.

The theory was revised accordingly:

  • Definition 1 and Proposition 1 now define hz:=ker⁡ℓz\mathfrak h_z:=\ker\ell_z directly. The injectivity step (ABNk) and the faithfulness and local-transitivity hypotheses are removed.

  • Theorem 1 is now a proposition with no novelty claimed, since local factorization holds for any complementary pair (JSx2). Reductivity is enforced by Ladj\mathcal L_{\rm adj} rather than assumed to hold.

  • Theorem 2 is restated: the foliation is by level sets of the classifier, and completeness holds for the local orbit exp⁡(hz)z\exp(\mathfrak h_z)z rather than all nearby points (eDXv).

  • Theorem 3 is renamed "Latent Displacement Bound" and no longer claims proximity to the data manifold (ABNk, JSx2). We strengthened it with two bounds that use the structure of the learned generators instead of their norms: a logarithmic-norm bound and an exact low-rank form. Measured over 10 runs (App. N.3), these are 3–16 orders of magnitude tighter than the worst-case bound and within a factor of 18–170 of the observed displacement. They capture the order of growth but are not numerically tight. Table 2 reports the displacement measured directly, which does not depend on any bound.

Several further points are clarified:

  • The classifier head is linear, so the decision geometry in latent space is affine (JSx2). We state this explicitly in the paper. The factual subspace still varies with zz through the group action.

  • Table 2 uses latent distance for every method; the leftover edit-distance text in App. M.1 is removed (ABNk).

  • GLIDE produces latent counterfactuals with node saliency rather than decoded graphs, which we state as a limitation (ABNk).

  • On BA-2Motifs, the top-5 nodes of the dominant factual generator recover the ground-truth motif exactly on 100% (cycle) and 92% (house) of test graphs (ABNk). We have moved this, as well as the explanatory subgraph visualizations to the main text.

  • The mechanism experiments are reframed as verifying, on held-out graphs, that trained models meet the theorems' hypotheses, rather than as independent evidence for the theorems (eDXv).

  • We replaced the supplementary code with the current version (JSx2).

We believe we have addressed all reviewer concerns, and are happy to answer any remaining questions throughout the discussion period.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.