acceptodds
Under review as a conference paper at ICLR 2027

Let persistence ShePHerD your sheaves

Abstract

Persistence-augmented graph neural networks enrich message passing with global topological information via persistent homology (PH). While some existing PH-based topological descriptors make the underlying graph filtration learnable, they fix a coefficient system. This imposes an unnecessary constraint that may limit the expressive power of PH-based descriptors and potentially hinder downstream predictive performance. To overcome this, we introduce ShePHerD as a model that learns the coefficient system with cellular sheaves, where we persist the resulting sheaf cohomology and enrich persistence events with the spectra of the sheaf Laplacian. Notably, we theoretically prove that integrating persistence sheaf cohomology into PH diagrams increases expressivity.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
persistent homology, cellular sheaves, GNN, topological neural networks

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_5TfTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a novel topological deep learning architecture combining two partially separate lines of works, persistent homology descriptors and learnable sheaf Laplacian, to achieve provably improved expressivity. In particular, the authors note that homology on different coefficients can be computed as cohomology over different cellular sheaves, opening an opportunity to combine these two approaches. Finally, the method is further improved by leveraging Sheaf Diffusion operators, typical in the sheaf neural networks literature.

Critical strengths

The manuscript is well written and motivated. The overall idea is interesting and, as far as I know, novel. The main construction is justified theoretically and empirically on synthetic graph data and some real banchmarks.

Critical weaknesses

While the paper is well written, I think a few details could be further clarified in the main manuscript.

In particular, my main doubt concerns Def. 3.4: do I understand correctly that this diagram is differentiable only with respect to the filtration times (via the birth and death indexes), but not with respect to the sheaf? How is the sheaf learnt then, when sheaf diffusion is not used? I see some argument in Apx. D.2 based on the spectral decorator (which is not explained much in the main paper), but it misses some details and arguments. Could the author clarify these aspects? In any case, I think differentiability is an important aspect that shouldn't be addressed only in the appendix, so I recommend the authors moving this discussion in the main manuscript.

Questions for authors

The example in the introduction showing the method can recover persisten homology on R\mathbb{R} is simple but useful. However, I wonder if the authors can include some more complex examples, e.g. can we construct a sheaf s.t. its cohomology recovers the persistent homology in Z2\mathbb{Z}_2?

Def 3.1: why is a G^\hat{G} needed in this definition?

End of page 4: what is EGE_G? Later is referred to as "eigenspace sum", does it refer to the space spanned by the selected eigenvectors?

Spectral connection above Eq. 3: why not sorting by eigenvalues? Also is LGL_G referring to the graph Laplacian or does it also support Sheaf Laplacians?

Eq. 4: do you have any interpretation or intuition for what this resulting orthogonal connection represent?

Def 3.4: what does spec>0spec_{>0} mean? this notation wasn't introduced earlier.

3 comments

Part 1/2

We are grateful for your careful reading of our manuscript and for the constructive suggestions regarding its presentation. Several of the points you raised were either discussed only in the appendix or left implicit in the main text, and we agree that they deserve clearer explanations. Following your feedback, we have expanded the discussion of differentiability in Section 3.2 and clarified the notation and interpretation of the spectral connection in Section 3.1. We address your questions below and would be happy to discuss any remaining concerns.

My main doubt concerns Def. 3.4: is this diagram differentiable only with respect to the filtration times, but not with respect to the sheaf? How is the sheaf learnt then, when sheaf diffusion is not used?

Thank you for raising this question. The sheaf is indeed learned through the spectral decorations, even when sheaf diffusion is not used. Your reading is correct regarding the discrete nature of persistence pairing, but there is an additional gradient path through the sheaf Laplacian that could have been better explained in the main manuscript.

More precisely, the selection of birth–death pairs is a discrete operation, and we do not differentiate through this combinatorial pairing. Once the persistence events are fixed, their filtration coordinates remain differentiable with respect to the learned filtration functions fVf_V and fEf_E.

The important distinction is that the connection parameters receive gradients independently through the spectral coordinate of Definition 3.4. For each persistence event, we construct the sheaf Laplacian from the learned transports Ru→vR_{u\to v} of Eq. (2). Since the matrix LF(R)=δ(R)⊤δ(R)L_{\mathcal F}(R)=\delta(R)^\top\delta(R) depends on these transports, changes to the connection affect its eigenvalues and consequently the output of the spectral readout.

For a simple eigenvalue λi\lambda_i with normalized eigenvector qiq_i, its derivative satisfies

dλi=qi⊤(dLF)qi.d\lambda_i=q_i^\top(dL_{\mathcal F})q_i.

This expression does not require differentiating the eigenvector. As in SpectRe, the spectral decoration retains the positive eigenvalues, including their multiplicities, alongside the persistence coordinates. ShePHerD uses the sheaf Laplacian in place of the ordinary graph Laplacian. Since eigenvalues are independent of the choice of eigenvector basis, the descriptor remains well-defined even when eigenvalues repeat. Wherever the selected spectrum is differentiable, its dependence on the learned transports provides a gradient path to the connection parameters, even without sheaf diffusion.

Following your recommendation, we have added a dedicated paragraph entitled "Differentiability" in Section 3.2, immediately following the definition of the descriptor (new text in blue). To complement the explanation, Appendix D.2 now includes a numerical gradient check.

The persistence events determine which features are represented, while the spectral decorations provide a differentiable route for learning the sheaf.

Can we construct a sheaf such that its cohomology recovers the persistent homology in Z2\mathbb{Z}_2?

Yes, and we appreciate the suggestion to make the relationship between ordinary persistence and sheaf cohomology more explicit.

A simple construction is the constant cellular sheaf F‾2\underline{\mathbb F}_2, where F2=Z2\mathbb F_2=\mathbb Z_2. We assign the vector space F2\mathbb F_2 to every vertex and edge, and take all restriction maps to be the identity.

Under this choice, the sheaf coboundary is precisely the graph incidence matrix computed modulo 22. Since −1=1-1=1 in F2\mathbb F_2, the signs associated with edge orientations coincide. Consequently, the resulting cellular sheaf cohomology is ordinary graph cohomology with coefficients in F2\mathbb F_2.

The same construction can be applied consistently at every level of a graph filtration. By the correspondence between homology and cohomology over a field, the resulting persistent cohomology has the same interval decomposition as ordinary persistent homology over F2\mathbb F_2.

This example follows from the constant-sheaf construction and the universal coefficient results discussed in Appendix A. It also helps clarify an important aspect of our contribution: while a constant sheaf recovers an ordinary coefficient system, our real-valued learned SO(d)SO(d) connections allow the restriction maps themselves to vary, thereby changing the resulting cohomological information.

Part 2/2

Def. 3.1: why is G^\widehat G needed in this definition?

Thank you for pointing this out. The auxiliary graph G^\widehat G is indeed unnecessary, since the connection can be defined directly on the input graph GG.

We have simplified Definition 3.1 accordingly, removing G^\widehat G and making the corresponding notation changes in the compatibility proofs of Propositions 3.2 and 3.3. The revisions are highlighted in the manuscript, with the previous notation crossed out in red and the replacement shown in blue.

The revised definition now states the construction directly on GG.

What is EGE_G? Does it refer to the space spanned by the selected eigenvectors?

Yes. The space EGE_G also provides the geometry underlying our spectral connection.

We regard R∣V∣\mathbb R^{|V|} as the space of real-valued functions defined on the vertices of the graph. The eigenvectors of the ordinary graph Laplacian LGL_G describe different patterns of variation across the graph: smaller eigenvalues correspond to smoother functions, whereas larger eigenvalues correspond to functions that vary more strongly across adjacent vertices.

The space EGE_G is the direct sum of the complete eigenspaces selected by our rule. If U∈R∣V∣×dU\in\mathbb R^{|V|\times d} is an orthonormal basis of this space, then

pv=U⊤ev∈Rdp_v=U^\top e_v\in\mathbb R^d

gives the coordinates of the orthogonal projection of the indicator vector of vertex vv onto EGE_G. In this way, the construction defines a spectral embedding v↦pvv\mapsto p_v, and the relative positions of these embedded vertices determine the orthogonal transports along the graph's edges.

We now define EGE_G in Section 3.1 (blue).

Why not sort by eigenvalues? Is LGL_G the graph Laplacian, or does it also support sheaf Laplacians?

Thank you for raising these two points. Our choice to order eigenspaces primarily by multiplicity is motivated by the computational cost of the resulting connection. Since the stalk dimension dd is the sum of the dimensions of the selected eigenspaces, choosing smaller eigenspaces first helps keep the sheaf Laplacian manageable.

More precisely, we order complete nonzero eigenspaces by increasing multiplicity, and use increasing eigenvalue to break ties. Importantly, we select entire eigenspaces rather than individual eigenvectors, because the latter are not uniquely determined within an eigenspace of repeated eigenvalue. This is also essential for the isomorphism compatibility established in Proposition 3.3.

The tie-breaking direction was incorrectly stated in the original Section 3.1, and we have corrected it in the revised manuscript. As discussed in our response to Reviewer KdH5, the corrected rule selects eigenvalue 1010 on SR25, consistently with Appendix B.7.

Here LG=D−AL_G=D-A is the ordinary graph Laplacian introduced in Section 2; the sheaf Laplacian is denoted separately by LFL_{\mathcal F}.

Eq. 4: do you have any interpretation or intuition for what this resulting orthogonal connection represents?

The geometric interpretation helps explain why the spectral connection can distinguish graphs even when individual transports have the same rotation angle.

The main idea is to associate an oriented plane with every edge using the spectral embeddings of its endpoints. For an edge (u,v)(u,v), the matrix Bu→vB_{u\to v} is a skew-symmetric generator of rotations in the two-dimensional plane spanned by pup_u and pvp_v.

Consequently, the transport

Ru→v=exp⁡(π2Bu→v)R_{u\to v}=\exp\left(\frac{\pi}{2}B_{u\to v}\right)

performs a rotation by 90∘90^\circ in this plane, while acting as the identity on its orthogonal complement.

An interesting consequence is that every individual edge transport has the same eigenvalues: ii, −i-i, and 11 with multiplicity d−2d-2. Therefore, the distinguishing information does not come from changing the rotation angle or the eigenvalues of individual transports.

Instead, it comes from how the rotation planes are oriented relative to one another across the graph. When these transports are composed along cycles, their relative orientations determine the resulting holonomy and influence the spectrum of the sheaf Laplacian. This provides a geometric explanation for why the construction can capture structural information that is absent from the ordinary graph Laplacian spectrum.

We have added this interpretation after Eq. (4) in Section 3.1 (blue).

Def. 3.4: what does spec⁡>0\operatorname{spec}_{>0} mean? This notation was not introduced earlier.

Thank you for noting the missing definition. The notation spec⁡>0(L)\operatorname{spec}_{>0}(L) denotes the multiset of strictly positive eigenvalues of LL, including their multiplicities.

We have made this explicit in Definition 3.4 (new text in blue), so the notation is introduced before being used in the descriptor.

We like to thank the reviewer once again for the valuable contribution. We firmly believe that the presentation quality of the paper is much better with these additions.

I thank the authors for the detailed answers, which clarified most of my doubts.

I recommend the authors to explicitly clarify the persistent homology path does do not provide gradients to learn the sheaf Laplacian, which instead come only through the spectral decorations (this seems still a bit implicit in the paragraph at the end of Sec 3.2).

Regarding this question:

Can we construct a sheaf such that its cohomology recovers the persistent homology in Z2?

Thanks for the answer, but this doesn't clarify how that can be represented by the Sheaf Laplacians learnt by your networks right? The stalks are in R^n, how can the network learn to produce this sheaf such that its cohomology recovers the persistent homology of Z2\mathbb{Z}_2? I think aspects about what sheaves and persistent homologies can be actually parameterized by the network need some further discussion in the paper.

I think the authors can clarify these points directly in the final manuscript, so I maintain my positive recommendation.

Reviewer_KdH5OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ShePHerD uses orthogonal cellular-sheaf connections to compute persistent sheaf cohomology and decorates its events with sheaf-Laplacian spectral information. It studies both learned feature-conditioned transports and a deterministic spectral construction; ShePHerD++ adds diffusion.

Critical strengths
  1. The coefficient system modification is an interesting deviation from RePHINE and SpectRe.

  2. Transport/Holonomy interpretation is straightforward.

  3. In terms of spectra, the construction distinguishes all 105 SR25 pairs and 311/400 BREC pairs.

Critical weaknesses
  1. Equation (2) considers the same MLP for both endpoint orderings, their outputs are subtracted from each other, arranged in a skew-symmetric matrix, and the result is exponentiated. When the endpoint embeddings are identical, the resulting transport is the identity. In the GIN(E) backbone setting, this takes place on symmetric regular examples with uniform features and matching incident edge attributes. Adding RWSE-16 and LapPE-8 (Appendix D.2) to GPS might help to break some ties, but the learned connection is not checked on the challenging examples with either backbone. Theorem 3.7 refers to a descriptor family, whereas SR25/BREC employ fixed spectral connection. I am therefore unsure what the learned connection actually separates.

  2. The performance improvement compared to SpectRe on PTC and IMDB-B is marginal – 63.5 vs. 63.4 and 77.6 vs. 77.0 (Table 2), both within the seed variation. One configuration is shared across all methods and datasets, not optimized for each dataset individually, and training stops at 60 epochs. For GPS, the MAE ZINC score is 0.246 on the above conditions, while around 0.070 on a published protocol. I would need more training time or better-tuned baselines to attribute any significance to the margins.

  3. No runtime/memory comparison to SpectRe is mentioned. Appendix D.2 discusses retries of eigendecomposition with a small structured diagonal perturbation near degenerate eigenspaces. How frequent are such retries, and what about gradients?

Questions for authors
  1. With either backbone, could the feature conditioned connection differentiate the graph pair in Theorem 3.7 or the hard BREC families without using the fixed spectral backbone?

  2. Can the improvements obtained by using real world data withstand further training or tuning with respect to the particular dataset? At what rate are the eigendecomposition retries?

  3. Was the eigenspace of eigenvalue 10 or 15 used for the SR25? Sections 3.1 and B.7 seem contradictory on that point.

5 comments

We sincerely thank you for the detailed review and the careful questions about our method. In particular, your observations helped us identify where the distinction between our two connection assignments was not sufficiently clear. We have revised the manuscript to clarify their respective roles, added a new expressivity result for the feature-conditioned connection, and conducted additional experiments addressing the experimental setup, computational cost, and numerical stability. We hope we addressed the majority of your concerns and would firmly believe the quality of our submission was increased.

Part 1/3

When the endpoint embeddings are identical, the resulting transport is the identity. [...] I am therefore unsure what the learned connection actually separates.

Thank you for highlighting the role of the feature-conditioned connection. To address your main concern, we have added a new pair of attributed graphs that is distinguished by the proposed learned SO(3)SO(3) connection itself. This establishes a strict expressivity result for the parameterization used by ShePHerD, alongside the structural separation demonstrated by the spectral assignment.

The construction is presented in Proposition B.4 and Figure 4 (Appendix B.9), with a reference added in Section 3.3. We consider two attributed graphs G1′G'_1 and G2′G'_2 that RePHINE, SpectRe, and edge-aware 11-WL cannot distinguish under any filtration induced by their attributes. Nevertheless, the feature-conditioned SO(3)SO(3) connection of Eq. (2) distinguishes them, as

dim⁡H0(G1′;Fθ)=1≠3=dim⁡H0(G2′;Fθ)\dim H^0(G'_1;\mathcal{F}_\theta)=1\neq 3=\dim H^0(G'_2;\mathcal{F}_\theta)

The difference comes from the transport accumulated around cycles, or holonomy. As established by Lemma B.2, the dimension of H0H^0 is determined by the vectors preserved by these transports. In one graph, the holonomy restricts the space of global sections to one dimension; in the other, all holonomies are trivial, leaving a three-dimensional space. The resulting difference in the persistence events and spectral decorations is illustrated in Figure 4. Importantly, the proof explicitly constructs the transports using the shared MLP parameterization of Eq. (2).

Indeed, Eq. (2) gives Ru→v=IR_{u\to v}=I whenever hu=hvh_u=h_v, because the two MLP evaluations cancel. The two assignments complement one another: the feature-conditioned connection exploits differences in endpoint representations, while the spectral connection constructs edge-dependent transports directly from graph structure, as used in BREC and SR25.

Thus, Theorem 3.7 establishes the expressivity of the general descriptor family, while Proposition B.4 demonstrates a strict separation using the actual feature-conditioned connection proposed in our model.

With either backbone, could the feature conditioned connection differentiate the graph pair in Theorem 3.7 or the hard BREC families without using the fixed spectral backbone?

Thank you for this question. On the uniform-feature pair of Theorem 3.7, GIN(E) gives identical endpoint representations, since every vertex has the same attributes and incident edge-attribute multiset; Eq. (2) therefore gives identity transport. GPS additionally uses structural and positional encodings, which can produce distinct endpoint representations beyond this edge-aware 11-WL argument. Theorem 3.7 demonstrates separation with a chosen orthogonal connection (−1-1 on blue edges, +1+1 on red edges), whereas the BREC and SR25 experiments use the spectral assignment. Proposition B.4 establishes the complementary strict separation using our feature-conditioned SO(3)SO(3) connection.

The new attributed pair in Proposition B.4 addresses the complementary question: it shows that the same feature-conditioned SO(3)SO(3) parameterization produces a nontrivial connection and strictly separates graphs that the persistence baselines and edge-aware 11-WL cannot distinguish. This is the learned-connection expressivity result highlighted in Figure 4.

Part 2/3

Was the eigenspace of eigenvalue 10 or 15 used for the SR25? Sections 3.1 and B.7 seem contradictory on that point.

Thank you for pointing out this inconsistency. The experiments use the eigenspace corresponding to eigenvalue 1010, with stalk dimension d=12d=12. This information was already given in Appendix B.7, but we agree that the selection rule in the main text did not communicate it correctly.

More precisely, the SR25 graphs have two nonzero Laplacian eigenvalues, 1010 and 1515, each with multiplicity 1212. Our selection rule first orders complete eigenspaces by increasing multiplicity and then breaks ties by increasing eigenvalue. Therefore, the selected eigenspace corresponds to eigenvalue 1010.

Unfortunately, Section 3.1 stated decreasing as opposed to increasing eigenvalue. We have corrected this wording in the revised manuscript, with the previous text crossed out in red and the corrected rule shown in blue. We appreciate you catching this discrepancy.

The performance improvement compared to SpectRe on PTC and IMDB-B is marginal [...] I would need more training time or better-tuned baselines to attribute any significance to the margins.

Thank you for this concern. Our experiments compare the descriptors under the same data splits, seeds, encoder settings, filtration networks, and training protocol. This shared setup makes the contribution of the proposed topological components directly comparable. We further evaluated the effect of a longer training budget for the standalone sheaf-network baselines.

The improvements over SpectRe on PTC_MR (0.10.1 percentage points) and IMDB-BINARY (0.60.6 points) are small relative to the reported seed variation, as stated in Section 4.1. The additional experiment below directly addresses the question of training budget.

To provide additional evidence regarding the training budget, we conducted a targeted longer-training experiment with NSD and CSNN, the two standalone sheaf neural networks included in our comparison. We extended their maximum training budget from 6060 to 200200 epochs while keeping the other settings unchanged. ShePHerD and ShePHerD++ retain their original 6060-epoch results.

The main finding is that the better of ShePHerD and ShePHerD++ at 6060 epochs remains ahead of both sheaf-network baselines trained for up to 200200 epochs on all six datasets examined. The comparison is reported in the new Appendix G, Table 10:

MethodMUTAGPTC_MRIMDB-BNCI1NCI109ZINC
NSD, 60 epochs62.157.152.070.269.00.530
NSD, 200 epochs62.157.152.070.769.40.504
CSNN, 60 epochs82.156.071.479.978.20.310
CSNN, 200 epochs76.856.072.478.478.80.268
ShePHerD, 60 epochs93.763.577.680.278.30.198
ShePHerD++, 60 epochs92.661.777.282.280.40.177

The longer budget does improve some baseline results, particularly on ZINC. However, these improvements do not close the observed gap with our methods in this comparison. For instance, the best ZINC MAE among the two extended baselines is 0.2680.268, compared with 0.1770.177 for ShePHerD++.

This additional experiment strengthens the comparison by showing that a longer training budget for NSD and CSNN does not close the observed performance gap.

Part 3/3

No runtime/memory comparison to SpectRe is mentioned.

We thank you for this suggestion. We have added training-time and peak-memory measurements against SpectRe in Appendix F (Tables 8 and 9), with a reference from the main experimental section.

The following table highlights the results on IMDB-BINARY and PROTEINS at matched float32 precision. Times are seconds per training epoch, and memory is peak GPU allocation.

MethodIMDB-B (s)PROTEINS (s)PROTEINS (MiB)
SpectRe67.12479.11650.4
Ours (d=1d=1, float32)68.16103.50503.7

Our spectral implementation achieves comparable runtime on IMDB-BINARY and is 4.63×4.63\times faster on PROTEINS, while using less peak GPU memory. In float64, it remains 3.64×3.64\times faster on PROTEINS. SpectRe is faster on smaller molecular graphs; the complete comparison is reported in Appendix F, Table 8.

The corresponding runtime and memory measurements are reported in Table 9. Since SpectRe performs eigendecomposition on the CPU, its reported peak GPU memory excludes the CPU eigensolver's working memory.

Appendix D.2 discusses retries of eigendecomposition with a small structured diagonal perturbation near degenerate eigenspaces. How frequent are such retries, and what about gradients?

Thank you for asking about this implementation detail. The eigensolver fallback is rarely needed in our experiments. When symmetric eigendecomposition fails to converge, we compute the singular values of the same sheaf Laplacian. Since this matrix is symmetric positive semidefinite, its singular values coincide with its eigenvalues. Appendix D.2 now describes this fallback and reports the numerical stability checks.

To quantify how often the fallback is needed, we instrumented 8,886,7408{,}886{,}740 decompositions on MUTAG, PTC_MR, IMDB-BINARY, and NCI1. The fallback was used only 1111 times, a rate of approximately 1.2×10−61.2\times10^{-6}, all on IMDB-BINARY. No nonfinite gradient entries were recorded.

We also examined the backward pass more directly. Since the spectral readout uses eigenvalues instead of eigenvectors, it avoids the eigengap denominators associated with eigenvector derivatives. We compared backpropagated gradients against central finite differences on C6C_6 and K5K_5, across 1212 configurations including repeated eigenvalues, and found agreement in every case.

We have added these measurements and the updated fallback description to Appendix D.2, and the explanation of how gradients reach the learned connection is now included in Section 3.2.

Reviewer_wCFQOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The core insight of this paper is elegant: existing persistent homology (PH)-based GNN descriptors rely on fixed coefficient systems (typically Z2 or R), imposing an unnecessary restriction on their representational capacity. The authors propose a principled generalization by using cellular sheaves to learn the coefficient system end-to-end. This connection to classical sheaf theory is mathematically well-founded. Specifically, the paper extends the framework of Hofer et al. (2020), which computes persistent (H_0) over learnable filtrations, by replacing fixed coefficients with learned cellular sheaves. This leads to the proposed Sheaf Persistent (co)Homology Diagrams (ShePHerD) model. The idea is that learning the coefficient system directly from data can enhance the expressive power and predictive performance of PH-based GNNs.

The paper also establishes several theoretical results, including invariance under graph isomorphisms and stronger expressive power compared with existing approaches such as RePHINE and SpectRe. These theoretical findings are further supported by illustrative graph examples and empirical evaluations on real-world datasets.

Critical strengths

Theoretical soundness and strong motivation.

Empirical evaluation is rigorous and comprehensive in terms of baselines.

Critical weaknesses

The major concern is two folds:

1, computationally the method, requiring eigendecomposition, of the sheaf laplacian could be quite expensive. And it could be even worse as the eigensolver needs to be differentiable (L1116, L1117). I recommend the authors report the training/inference time and discuss this limitation. It does not really lower the scientific value of the paper in my opinion.

2, the results on BREC and SR25 are strong and interesting. But the performance on real world datasets shows marginal gains, perhaps because the graphs are generally of moderate sizes.

Questions for authors

See weakness regarding reporting running time.

3 comments

Part 1/2

We thank you for your constructive review and for recognizing the expressivity results on BREC and SR25. We particularly appreciate your suggestion to examine the computational cost more carefully, as this is an important practical consideration for methods involving spectral decompositions. Following your feedback, we have added new runtime and memory experiments, together with a detailed computational-cost comparison in Appendix F.

I recommend the authors report the training/inference time and discuss this limitation. It does not really lower the scientific value of the paper in my opinion.

Thank you for this suggestion. We have added training-time and peak-memory measurements that quantify the cost of ShePHerD and isolate the incremental expense of learning the connection at a fixed stalk dimension. The results are reported in Appendix F (Tables 8 and 9), with a reference from Section 4.1.

As you noted, the descriptor requires eigendecomposition of the sheaf Laplacian associated with the components of the persistence events. The matrix has dimension d∣V∣d|V|, rather than ∣V∣|V| for the ordinary graph Laplacian, so the computational cost increases with the stalk dimension. In our learned experiments, we use d=3d=3 to allow noncommuting rotations while keeping the operator relatively small.

To better understand the practical implications, we performed two complementary comparisons. First, we compared our spectral implementation with the released SpectRe kernel under a matched eigenproblem. Second, we measured the additional cost of the learned connection within our own implementation.

Comparison with the released SpectRe implementation. SpectRe computes the ordinary graph-Laplacian spectrum using a C++ CPU kernel in float32, while our layer computes spectra using PyTorch on the GPU. We report our layer in both float32 and float64, the latter being the precision used in our predictive experiments.

At matched float32 precision, our implementation is slower on small molecular graphs, reaches approximately the same speed on IMDB-BINARY, and becomes faster on PROTEINS. The crossover illustrates how runtime depends on the graph and eigensolver implementation as well as the spectral problem size.

The measurements are reported in Table 8 of Appendix F. All experiments use the same NVIDIA RTX 2080 Ti, and times are measured in seconds per training epoch.

DatasetMean nodesSpectReOurs (float32)Ours (float64)
PTC_MR14.32.0510.3416.94
MUTAG17.91.338.1213.30
IMDB-BINARY19.867.1268.16136.97
PROTEINS39.1479.11103.50131.56

On PROTEINS, our implementation is 4.63×4.63\times faster at matched float32 precision and 3.64×3.64\times faster in float64. Conversely, on MUTAG and PTC_MR the released SpectRe kernel remains substantially faster. The results show a crossover in runtime rather than a uniform advantage.

We also measured peak GPU memory, reported in the same table. These measurements require some care when interpreting them: SpectRe performs eigendecomposition on the CPU, so its GPU peak does not account for the CPU eigensolver's working memory.

Part 2/2

The second comparison isolates what changes when we replace the identity connection with the learned SO(3)SO(3) connection, keeping the implementation and stalk dimension fixed.

From Table 9 (Appendix F):

MethodMUTAGPTC_MRNCI1
Identity connection (d=3d=3)48.2463.36461.62
ShePHerD (d=3d=3)57.7575.12577.18
ShePHerD++ (d=3d=3)52.0277.24520.64

Relative to the identity connection, ShePHerD has ratios of 1.20×1.20\times, 1.19×1.19\times, and 1.25×1.25\times, respectively. The released SpectRe kernel is faster in this setting, as it solves a smaller scalar eigenproblem using a different implementation.

However, the matched d=1d=1 comparison in Table 8 also reveals an important advantage of our spectral implementation: on PROTEINS, it is 4.63×4.63\times faster than SpectRe in float32 and 3.64×3.64\times faster in float64, with lower peak GPU memory in float32. On IMDB-BINARY, the float32 runtimes are nearly identical, while SpectRe remains faster on MUTAG and PTC_MR.

The larger sheaf Laplacian introduces additional work, but learning the connection itself accounts for only 1919\\% additional training time relative to the identity-connection implementation at d=3d=3. Appendix F discusses this distinction.

We also clarify in Section 3.2 and Appendix D.2 that the backward pass uses eigenvalue information without requiring derivatives of eigenvectors, and report the numerical stability checks conducted for this computation.

The results on BREC and SR25 are strong and interesting. But the performance on real world datasets shows marginal gains, perhaps because the graphs are generally of moderate sizes.

Thank you for this observation. We agree that stronger graph distinguishability does not necessarily translate into equally large improvements on every predictive task. However, ShePHerD and ShePHerD++ achieve the highest mean performance across all seven real-world benchmarks, with several improvements that go beyond the marginal differences highlighted in the review.

In particular, ShePHerD improves over the strongest baseline by 2.12.1 percentage points on MUTAG and 0.90.9 on MOLHIV, while ShePHerD++ improves by 1.51.5 points on NCI1 and 2.02.0 on NCI109. On ZINC, ShePHerD++ reduces the test MAE from 0.2130.213 to 0.1770.177, a relative reduction of approximately 16.916.9\\%. We also acknowledge that the improvements over SpectRe on PTC_MR (0.10.1 points) and IMDB-BINARY (0.60.6 points) are within the reported seed variation, as discussed in Section 4.1.

Importantly, we have also strengthened our theoretical results with a new expressivity construction for the actual feature-conditioned connection used in ShePHerD. In Proposition B.4 and Figure 4 (Appendix B.9), we construct two attributed graphs that RePHINE, SpectRe, and edge-aware 11-WL cannot distinguish under any attribute-induced filtration. Nevertheless, our proposed SO(3)SO(3) connection, realized through the shared MLP parameterization of Eq. (2), distinguishes them:

\neq 3=\dim H^0(G'_2;\mathcal{F}_\theta).$$ This result is particularly relevant to real-world graph learning because the separation uses the feature-conditioned connection employed in our predictive experiments, rather than the spectral connection used on BREC and SR25. While the construction is synthetic, it establishes that the proposed learnable parameterization itself provides additional distinguishing power on attributed graphs. We therefore consider the real-world and expressivity experiments complementary. The predictive benchmarks evaluate the benefit of learning the coefficient system under a shared training protocol, while the expressivity results establish distinctions unavailable to the persistence baselines. In particular, ShePHerD distinguishes $311/400$ BREC pairs compared with $216/400$ for SpectRe, and the spectral connection distinguishes all $105$ SR25 pairs, whereas RePHINE and SpectRe distinguish none. Proposition B.4 further extends this evidence to our feature-conditioned connection. We have also added a targeted longer-training comparison in Appendix G. NSD and CSNN were trained for up to $200$ epochs instead of $60$, while our methods retained their original $60$-epoch results. **The better of ShePHerD and ShePHerD++ remains ahead of both baselines across all six datasets examined**, despite the shorter training budget. We appreciate your suggestions, which helped us clarify the computational trade-offs and strengthen the evidence for the benefits of learning the coefficient system.

Thank you for the rebuttal. It addressed my concerns.

Comments on the paper

1 comment

We would like to thank the reviewers, the Area Chair, and the Program Chairs for their time and consideration of our work. We are particularly grateful for the constructive feedback, which helped us strengthen both the theoretical and experimental contributions of our paper.

During the rebuttal period, we have made several additions to the manuscript, among which we would like to highlight three:

  • Expressivity of the learned connection (Proposition B.4 and Figure 4). We have shown that ShePHerD with the proposed feature-conditioned SO(3)SO(3) connection can distinguish a pair of attributed graphs that RePHINE, SpectRe, and edge-aware 11-WL cannot distinguish. This new result complements Theorem 3.7 by establishing a strict separation using the same learnable parameterization employed in our real-world experiments.

  • Stability of ShePHerD (Theorem 3.9 and Appendix E). We have established a local stability theorem for ShePHerD with respect to perturbations of both the filtration functions and the sheaf connection. This extends the stability guarantees of SpectRe to learnable coefficient systems under injective filtration functions and locally constant coboundary ranks.

  • Differentiability of ShePHerD (Section 3.2 and Appendix D.2). We have provided a more detailed mathematical explanation of how the coefficient system is learned directly through the spectral decorations of our persistence descriptor, without requiring sheaf diffusion or differentiating through eigenvectors.

We are glad that Reviewers KdH5 and wCFQ have indicated that our rebuttal addressed their concerns. We believe that the new theoretical results and additional experiments have strengthened the paper as a whole, and we kindly ask all reviewers to consider whether these improvements warrant a higher score. We would also be more than happy to address any remaining questions.

Once again, we sincerely thank the reviewers, the AC, and the PCs for their consideration of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.