acceptodds
Under review as a conference paper at ICLR 2027

TWINS-GCN: SPECTRAL CONJUGATE FOR SPECTRAL GRAPH CONVOLUTIONAL NETWORKS

Abstract

Graph convolutional networks propagate information by repeated local aggregation through a graph shift operator; i.e., a -layer network reaches hops neighborhood. On the one hand, such spreading can lead to oversmoothing. On the other hand, long-range dependencies demand the depth. Transporting information on long distances and without attenuation requires the shift to distinguish a direction of flow, which a symmetric operator cannot perform but a directed one can fulfill. A natural way to extract pure-directionality is to take the skew-symmetric part of the shift operator through the Cartesian split, which, however, generally does not commute with the shift itself, meaning that the filters built on it are not shift-invariant. We instead use the spectral conjugate; i.e., the image of the operator under , which commutes with the shift and splits it into a dissipative and a non-dissipative part. Two filter families follow: a sum filter, whose non-dissipative component transports signal without energy loss, and a ratio filter, ratio in the pair of components rather than polynomial in the shift. Both arise from non-holomorphic kernels, placing them outside the class any classical spectral convolution can reach. On the directed cycle, the ratio filter becomes an IIR filter with global impulse response, for which we prove a long-range reach gap against every degree- polynomial filter. Chebyshev reparameterization gives stable vertex-domain layers with real coefficients, yielding TwinS-GCN, which solves graph transfer tasks at reduced depth and is competitive with state-of-the-art graph convolutional networks on node classification benchmarks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
spectral graph convolution, directed graphs, graph signal processing, graph neural networks

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_Yr9bOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

At a high level, this paper introduces TWINS-GCN, a spectral framework for convolution on directed graphs. Its key move is to use the spectral conjugate of the graph operator to split it into commuting dissipative and asymmetric parts. On top of that split, the authors build sum and ratio filters, derive vertex-domain forms that also cover non-diagonalizable operators, and give a Chebyshev-based parameterization for learning real-valued layers. The most striking theoretical result is that, on a directed cycle, the ratio filter behaves like an IIR filter and can reach beyond the KK-hop limit of degree-KK polynomial filters. The authors validate the approach on synthetic long-range transfer tasks and five directed node-classification benchmarks, where TWINS is competitive overall but does not win consistently.

Critical strengths
  1. I think the operator construction is principled and addresses a real limitation. The paper makes a clear case for why the usual Cartesian split can fail to commute with a non-normal graph operator, then uses the spectral conjugate to build a commuting alternative. Extending the construction to non-diagonalizable operators also gives the work more substance than a straightforward architecture tweak. I personally believe this is the paper’s strongest contribution: it offers a useful way to think about attenuation and directional transport within one spectral framework.

  2. The ratio filter gives a crisp, useful long-range result. The IIR interpretation on directed cycles, together with the reach analysis, is both intuitive and technically interesting. The transfer task is also well matched to that claim: TWINS variants can move information across small graphs with fewer layers than several polynomial baselines. That gives the theory a concrete, if still limited, empirical anchor.

Critical weaknesses
  1. The general non-normality claims need a tighter link to the quantities the paper says it preserves. The fact that eigenmode responses have unit modulus does not, by itself, mean that signal norms or Dirichlet energy are preserved on a non-normal graph: its eigenvectors may be highly non-orthogonal, and the operator can still amplify or shrink signals. The energy argument is clearest on the normal directed cycle; the broader claim about retaining feature distinctiveness is less firmly supported. I’m not saying the underlying theory is wrong, but the paper should be more careful about what its guarantees do—and do not—imply for general graphs. Likewise, Chebyshev reparameterization may help with coefficient conditioning, but it does not by itself establish numerical stability for a non-normal matrix polynomial.

  2. The practical case is not yet strong enough, given the cost and the scope of the experiments. Computing the spectral conjugate requires a Schur-based procedure with O(N3)\mathcal{O}(N^{3}) cost, and the ratio branch may involve a dense quotient operator. The largest benchmark has about 5,200 nodes, and the reported build time for the ratio model on Squirrel-F is around 58.5 seconds per graph; there are no memory figures or scaling curves to show what happens beyond that range. The benchmark results are also mixed: for example, TWINS (R) reaches about 40.15% on Squirrel-F, versus 45.13% for FaberNet. Since some baseline results come from prior work and the tuning and validation protocols are not fully matched, it’s hard to tell how much of the overall advantage is robust.

Questions for authors
  1. For general graphs, do you explicitly materialize Lϕ=L↑(L°)DL^{\phi}=L^{\uparrow}(L^{\degree})^{D} to apply the ratio filter, or is there an implicit implementation? Could you report its memory use and end-to-end build, training, and inference time as graph size and edge count grow?

  2. What exactly is guaranteed by the “non-attenuating” argument on a non-normal graph: unit-modulus responses in an eigenbasis, preservation of a signal norm, or preservation of the paper’s Dirichlet energy? Do you have bounds or experiments showing how the effect changes with eigenvector conditioning or other measures of non-normality?

  3. Could you clarify how the main benchmark comparison changes under a matched evaluation protocol—for example, using the same data splits and comparable hyperparameter-search budgets for TWINS and the strongest baselines?

(Very minor) A comparison with directed adaptations of rational filters, such as ARMA- or Cayley-style methods, would help separate the benefit of the commuting split from the benefit of using a higher-reach filter family.

7 comments

We are grateful to the reviewer for the detailed and constructive review. We address Weakness 1 and Questions 2 and 4 (very minor) below. The points on cost, scaling and the matched evaluation protocol (Weakness 2 and Questions 1 and 3) will follow in a separate comment, together with the results of the additional experiments we are currently running.


R3.1 Non-attenuating guarantee and Dirichlet energy

We agree with the reviewer that the text would benefit from clarifying the object of non-attenuation. In Section 3.3.1, we first discuss the non-attenuation of Dirichlet energy in CN\mathcal{C}_N since beyond this case, to our knowledge, no immediate link can be drawn between the Dirichlet energy computed over a symmetrized edge set and the directed graph Fourier coefficients x^\hat{\mathbf{x}}. Nonetheless, we take a step further in the observation of equation 18, where non-attenuation of Dirichlet energy amounts to non-attenuation of the x^\hat{\mathbf{x}} (the coefficients on the Fourier eigenbasis). The guarantee we provide is then on the non-attenuation of these Fourier coefficients, which is exact for diagonalizable L, and bounded from below for defective L (bounds given in Appendix C.5). In short, we provide guarantees where analytically possible, and arguments for the preservation of Dirichlet energy in more general graphs through experiments. This reasoning is made explicit in the revised version (Section 3.3.1). Following the reviewer’s suggestion, we now also discuss in the text (Section 3.3.1) the effect on signal norm, providing upper and lower bounds in terms of the condition number of the generalized eigenvectors (Appendix C.5).


R3.2 Chebyshev reparametrization

We thank the reviewer for the remark. We agree that the “stability” wording is inaccurate as the Chebyshev reparametrization does not make the matrix polynomial numerically stable in general. What it improves is the conditioning of the filter coefficients, which is known to be poor in the monomial basis (Boyd, 2001). Therefore, following the reviewer’s suggestion, we no longer use “stable” throughout the manuscript and instead appropriately justify the reparametrization in Section 3.1.


R3.3 Comparison with rational filters

Although directed variants of ARMA and Cayley filters have been developed in directed graph signal processing, we were unable to find a node-level directed GCN implementation of them that we could adapt and tune within the discussion period. Conceptually, rational filters share the higher reach of IIR filters but remain holomorphic (Appendix C.4). Hence, as discussed in in our reply to Reviewer 1 (R1.2), ARMA or Cayley filters cannot set the responses to dissipation and transport independently. The ratio filter, in contrast, is built on the quotient L↑(L∘)D\mathbf{L}^{\uparrow}(\mathbf{L}^{\circ})^D, whose spectrum is purely imaginary, so its far reach does not have to come with dissipation. Beyond the directed cycle, where the ratio filter is itself rational in the shift, a low-order rational filter (ARMA) generically cannot achieve transport without dissipation. We now discuss this in Section 3.3.2 and in more details in Appendix C.8.

AuthorsOpenReview ↗

Follow up on additional experiments

R3.4 Filter implementation and cost

We explicitly compute the operators (Z~\tilde{\mathbf{Z}}, L~↑\tilde{\mathbf{L}}^{\uparrow} and L~∘\tilde{\mathbf{L}}^{\circ}), which may be dense, once per graph before training using a single Schur decomposition. The build therefore costs O(N3)O(N^3) time and O(N2)O(N^2) memory, but it is paid only once and the matrices are reused for training and inference. Following the reviewer's suggestion, we now provide plots of build, per-epoch training, inference time and memory usage as a function of graph size for TwinS-GCN and other baselines models, in Appendix E. In these experiments, we consider increasingly large directed Erdős–Rényi graphs with fixed edge density. As expected, the build cost of TwinS-GCN is the highest across models. However, training and inference time, as well as memory usage, are comparable to those of models such as MagNet and FaberNet in this setting. In particular ADR-GNN is magnitudes more expensive for inference time and memory relative to all other models including TwinS. The cost of Twins-GCN is thus concentrated in a single precomputation, after which it trains at a cost comparable to other spectral GCNs.

This matches the setting our method targets, that is, small and medium directed networks where interpretability matters, such as temperature networks (Appendix F) or parcellated brain graphs (10210^2 to 10310^3 regions). Spectral GNNs that use the full spectrum share this cost. Specformer (Bo et al., 2023), for instance, relies on an O(N3)O(N^3) eigendecomposition and turns to a truncated one on large graphs. Unlike Specformer, which works with symmetric Laplacians, we cannot rely on an eigendecomposition, simply because a directed Laplacian may be non-diagonalizable and its eigenvectors can be very ill-conditioned. We therefore build on the Schur decomposition, which always exists and is computed stably. A truncated Schur decomposition, for instance with Krylov-Schur method would be the analogous route for TwinS-GCN on larger graphs, which we discuss in the limitations (Section 6).


R3.5 Matched protocol

As announced in our previous reply, we report results for all directed baselines (MagNet, Dir-GNN, FaberNet, ADR-GNN, AdvDIFFormer), which are now run under the same protocol as TwinS-GCN:

  • the configurations are selected on mean validation accuracy over the same ten splits
  • the same number of trials as TwinS-GCN with 72 on the WebKB graphs and 32 on Chameleon-F and Squirrel-F. Previously the number of trials were larger for the baselines.
  • the same shared space for all models. Same K and depth now applies to all models.

Only GCN and GAT, and MLP on Chameleon-F/Squirrel-F are taken from literature (non competitive baselines). We report the newest numbers in the following Table.

ModelChameleon-FSquirrel-FTexasWisconsinCornellMean rank
MLP36.73 ± 4.7136.55 ± 1.8282.97 ± 5.5085.56 ± 4.4676.22 ± 3.916.80
GCN40.89 ± 4.1239.47 ± 1.4752.16 ± 5.1648.92 ± 3.0652.70 ± 5.3010.00
GAT39.21 ± 3.0835.62 ± 2.0658.38 ± 6.6349.41 ± 4.0954.32 ± 5.0510.20
MagNet40.25 ± 3.8441.27 ± 1.7581.44 ± 4.8283.53 ± 4.2275.05 ± 2.856.50
Dir-GNN43.80 ± 4.4343.48 ± 1.6569.91 ± 6.1171.70 ± 6.4459.46 ± 5.496.60
FaberNet44.51 ± 3.6445.81 ± 2.1983.96 ± 5.3681.76 ± 3.5175.05 ± 4.823.90
AdvDIFFormer42.49 ± 2.9439.83 ± 1.8883.51 ± 4.9079.35 ± 5.2171.53 ± 4.236.60
ADR-GNN45.30 ± 3.8141.34 ± 1.8082.34 ± 4.4085.95 ± 4.1274.14 ± 4.114.40
TwinS-GCN (S)42.33 ± 2.8039.61 ± 1.5584.59 ± 3.4387.12 ± 4.1375.14 ± 3.154.40
TwinS-GCN (R)44.24 ± 3.2340.15 ± 1.8384.05 ± 3.9187.39 ± 3.9876.76 ± 3.012.60
TwinS-GCN (C)41.64 ± 3.5940.10 ± 1.5885.41 ± 4.7186.80 ± 3.7975.68 ± 4.014.00
*Mean-rank in the above table are computed for the listed 11 models.

Overall, the results of the matched protocols lead to similar ranking as in the initial submission, with TwinS-GCN (R) leading (2.6, against 2.8 in the original submission), FaberNet overall second, and the gap on Squirrel-F remains but the overall comparison is unchanged. TwinS-GCN (C) ranks 3rd, and TwinS-GCN (S) ranks 4th place along ADR-GNN. The revised Table 1 reports these results, now with the two newer baselines (FSpecGNN and DHGNN) discussed in our reply to Reviewer 1.

AuthorsOpenReview ↗

Summarizing changes

We thank the reviewer once more for the careful and constructive review, and for recognizing the filters and operator construction and the long-range results as the main strengths of the paper. The reviewer’s comments led to concrete improvements in the revision:

  • the non-attenuation guarantee is now stated precisely, with a bound on the signal norm in terms of the conditioning of the generalized eigenvectors (Section 3.3.1, Appendix C.5)
  • we now justify Chebyshev reparametrization, stating the conditioning of the filter coefficients rather than claiming stability (Section 3.1);
  • we provide a comparison with rational filters in Section 3.3.2 and Appendix C.8
  • we measure the cost of the method and state its scope (Appendix E and, conclusion and limitations)
  • all directed baselines are now compared under a matched protocol (Table 1)

We believe these changes address the concerns raised, and we would be very grateful if the reviewer would consider revisiting their assessment given these changes. We remain available for any further questions during the discussion period.


References

(Specformer) D. Bo, C. Shi, L. Wang, and R. Liao, “Specformer: Spectral graph neural networks meet transformers,” in The Eleventh International Conference on Learning Representations, 2023.

Reviewer_x27iOpenReview ↗
Rating
2
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces spectral conjugate for graph convolutional networks.

Critical strengths

The topic is interesting and timely, namely the idea of introducing new concepts to address issues within the operator in graph convolutional networks.

Critical weaknesses

The paper is very difficult to read and understand. It explores and introduces too many concepts. It is not clear if these concepts are necessary to address the raised problem.

I have tried to check the derivations and the obtained results, but I did not get the same results. I give a straightforward example in the following: The spectral conjugate is central in the paper. In (4) and in (63), it is defined as A♯=PΛ‾P−1A^\sharp = P \overline{\Lambda} P^{-1} for a matrix A with the canonical form A=P(Λ+N)P−1A = P (\Lambda + N) P^{-1}. It seems to me that this is not correct, because the nilpotent N has disappeared without any reason. I think that it should depend on the matrix N, and thus the spectral conjugate should be A♯=P(Lambda‾+N)P−1A^\sharp = P(\overline{Lambda} + N)P^{-1}. If this is not correct, then most of the derivations in the paper have major issues.

It is not clear why we need to use the Drazin inverse. Although it is used in Theorem 2, it seems to me that it complexifies the formulations. Moreover, it is only used in that theorem. Although it is defined in Appendix C.2, it is not used in practice, but via the complex Schur decomposition, which is already computed for the spectral conjugate.

There are theorems and a corollary. But it is not clear what is important in these results, and what is used next. For example, most of Theorem 1 follows from Nevanlinna (2018). It would have been relevant to present only the main theoretical results that are relevant for this work, and explain clearly why they are relevant for this work.

It is difficult to understand the TwinS architecture, as well as its different variants. It would have been relevant to provide some illustrations.

I acknowledge the author for their statement on the use of generative AI tools to assist with the writing of proofs. However, properties and theorems are introduced without knowing why. Moreover, the proofs and properties also explore other concepts brought by AI from different contexts and frameworks, making it more difficult for the readers to follow and understand.

Questions for authors

none

3 comments

We thank the reviewer for their time in the review and rechecking the derivations. We reply to the reviewer’s feedback in the following points.


R2.1 Nilpotent vanishing in spectral conjugate

In the non-holomorphic calculus introduced by Nevanlinna (2018), the action of a non-holomorphic kernel on a Jordan block Jn\mathbf{J}_n (Lemma 7 in Appendix B.2) reads

h(Jn)=∑k=0m−1∂kh(λn)Nkh(J_n) = \sum_{k=0}^{m-1}\frac{\partial^k h(\lambda_n)}{N^k}

Now replacing hh by τ\tau (conjugate kernel), we observe that ∂kτ=0\partial^k \tau = 0 for k>=1k>=1, yielding τ(Jn)=τ(λn)I\tau(\mathbf{J}_n) = \tau(\lambda_n)\mathbf{I} and thus L♯=τ(L)=Pτ(J)P−1=PΛˉP−1\mathbf{L}^{\sharp} = \tau(\mathbf{L}) = \mathbf{P}\tau(\mathbf{J}) \mathbf{P}^{-1} = \mathbf{P}\bar{\mathbf{\Lambda}} \mathbf{P}^{-1}, with J\mathbf{J} being the Jordan blocks. We agree with the reviewer that this was previously not clarified, we have now made it explicit in Appendix C.1, which is referred to right after Definition 2.1. Following the reviewer’s concern, we have also taken this as an opportunity to confirm the subsequent proofs and derivations.


R2.2 Drazin inverse

We introduce the Drazin inverse because it guarantees properties that are necessary for the implementation of the ratio filter, which the more common Moore-Penrose pseudoinverse does not satisfy. Specifically, the ratio filter requires

  • inverting the eigenvalues of L∘\mathbf{L}^\circ to match the definition of the ratio kernel (Section 2.4) and
  • for features to be analyzed in the same basis (linear shift-invariant property) the chosen inverse must commute with L\mathbf{L}.

Drazin inverse allows us to invert non-zero eigenvalues and maps any zero eigenvalue to zero, while the Moore-Penrose pseudoinverse inverts the singular values. For normal matrices they can coincide, however for more general matrices they do not. Additionally, the Drazin inverse (L∘)D(\mathbf{L}^{\circ})^D commutes with L\mathbf{L} while the Moore-Penrose alternative does not. As discussed by the reviewer, to compute the Drazin inverse we reuse the Schur decomposition already computed for the spectral conjugate. However, obtaining the full Drazin inverse, as we do, requires not only inverting the appropriate diagonal entries but also adjusting off-diagonals (Appendix C.2). Therefore, we believe its introduction is justified, as it provides the appropriate mathematical operation and its computation can leverage previously computed quantities. In the revised manuscript, we now provide a more detailed motivation for this choice in Section 2.4, where the matrix is introduced.


R2.3 Clarity, readability and relevance of concepts

We agree that the paper would benefit from clearer justification of its theoretical results. In the revision, each proposition, theorem and corollary is now stated along a short rationale, making sure to convey why it is needed and how it relates to the rest of the paper. Results taken from Nevanlinna (2018) are now always specified alongside its citation, separated from our contribution. To address the reviewer's concerns about clarity, we have also revised the supplementary material in Appendix C so that each subsection now begins or ends with a key takeaway that places its content in the broader context of the paper. We have also ensured that each subsection is explicitly referenced in the main text, with the exception of the former Appendix C.6, which has been removed in line with the reviewer's broader suggestions. Also, following the reviewer’s suggestion, we introduced Figure 1, which shows a schematic of the architectures, giving an overview for the construction of each proposed model. We believe that the figure provides valuable additional support by illustrating how the filters are constructed using the spectral conjugate, how they are integrated into the model architecture, and how they are implemented in practice, improving the overall clarity and accessibility of the work. As concerns the use of artificial intelligence, we would like to reassure the reviewer that while AI has been leveraged to help in proofs, the properties and concepts introduced were written and selected by us to provide what we believe is useful background information.


We would like to again thank the reviewer for their feedback which we believe has contributed significantly to improving the quality and readability of the manuscript. With the changes implemented, we would be very grateful if the reviewer would consider revisiting their assessment. We would be happy to clarify any remaining points during the discussion.

Reviewer_x27iOpenReview ↗

Rebuttal

I would like to thank the authors for their feedback. Following the reviews, they seek to address the raised issues one by one. While some modifications are of interest, such as including the illustration, others remain open questions, such as the relevance of introducing and using the Drazin inverse, which is still not well motivated to my eyes (only very specific to Theorem 2, but not explored in detail and without any interest elsewhere). Overall, the paper remains globally difficult to read and understand, although the requested localized modifications and the added illustration.

We thank the reviewer for their follow-up and for acknowledging the added illustration.

We understand that the role of the Drazin inverse may have appeared limited to Theorem 2. Building on our previous reply, we would like to further clarify that the Drazin inverse’ role goes beyond Theorem 2, as it is the unique operator satisfying the properties required for the ratio filter described in our previous reply (Campbell & Meyer, 2009). This is why we introduce it, rather than one choice among several, it is the operation that makes the ratio filter well defined. As the building block of this filter, it is also explicitly computed throughout all our experiments.

On readability, as described in our previous reply, each result is now introduced by a short statement of why it is needed and where it is used, and we added the illustration of the construction. We would value the reviewer’s guidance to make sure our remaining revisions target what matters most, could the reviewer point us to the one or two places where the reading is least clear? We would be glad to revise them before the end of the discussion period.

We are again truly grateful to the reviewer for their time and feedback.


References

Stephen L. Campbell and Carl D. Meyer. Generalized Inverses of Linear Transformations. Classics in Applied Mathematics. SIAM, 2009.

Reviewer_1fKxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces TWINS-GCN that solves graph transfer tasks at reduced depth and is competitive with state-of-the-art graph convolutional networks on node classification benchmarks.

Critical strengths

This paper proposed a new idea of using two types of filters including sum filter and ratio filter. By applying Twins-GCN, most dataset suggests better performance in classification tasks than SOTA method.

Critical weaknesses
  • The math analysis and result is solid, but the idea is an incremental work that focus on filter design of GCN.
  • The tasks only includes graph transfer and node classification. Besides baselines like GCN, GAT, MLP, 4-5 other models like AdvDIFFormer (published in 2023) and ADR-GNN (published in 2024) were used to compare with TWINS-GCN. I would like to see the comparison with other newer models from more recent years.
Questions for authors

Why did you choose those two filter families instead of others?

3 comments

We thank the reviewer for finding the analysis and result solid. We address the first two concerns below. The points on newer baselines and task choice (R1.3) will follow in a separate comment, together with the results of the additional experiments we are currently running.


R1.1 On incrementality

As for previous works on Faber-, Cayley- and Mag-Net, the novelty in spectral directed GCN often lies in the design of the graph filters. While, as pointed out by the reviewer, a significant contribution of this work lies in GCN filter design, we believe that this is not merely an incremental advance with respect to the existing literature but rather a new formulation with important practical consequences in terms of performance and interpretability. For clarity, we highlight the biggest novelty points below.

  1. Our implementation fundamentally differs from existing spectral GCNs, which parametrize filters as a polynomial or rational functions of the shift operator, thereby only expressing holomorphic kernels (Appendix C.4). In contrast, our proposed sum and ratio kernels are non-holomorphic (Theorem 2) and thus fall outside the functional class considered by previous spectral GCNs, while retaining interpretable spectral responses. Consequently, we introduce a new set of filters and implement them for the first time for directed spectral GCN. The differences with respect to previous formulations lie primarily in the fact that the holomorphic kernels underlying their filter architectures cannot depend exclusively on either ℜ(λ)\Re(\lambda) or ℑ(λ)\Im(\lambda), and thus cannot independently control dissipation (ℜ(λ)\Re(\lambda)) and transport (ℑ(λ)\Im(\lambda)), as our proposed sum filter is designed to achieve. While it is true that, on a fixed graph with N nodes, our proposed graph filters can also be expressed as polynomials in L\mathbf{L}, — represented within the operator Krylov space generated by L\mathbf{L}— they generally require polynomials of degree up to N – 1, with ill-conditioned coefficients (Nevanlinna (2018) & Appendix C.10). As spectral GCN usually have depth (degree) widely below the graph size, in practice the graph filters we propose are out of reach for existing spectral GCNs.
  2. We also differ from directed spatial models such as Dir-GNN that combine L\mathbf{L} and LT\mathbf{L}^T, which do not commute with L\mathbf{L} on non-normal graphs (Prop. 1), so their filters are not jointly interpretable in the same spectral domain. With all of that, we believe that we propose an elegant yet novel contribution to spectral directed GCNs. We have now revised the manuscript to better convey this rationale.

R1.2 Why these two families?

As discussed in R1.1, we believe that the sum filter is a natural way to give independent control of ℜ(λ)\Re(\lambda) (dissipation) and ℑ(λ)\Im(\lambda) (transport). However, this also means that the sum kernel is composed of a function that only sees the real part and another function that only sees the imaginary part of the spectrum. This is why we also propose the ratio kernel (and the derived ratio filter) which accounts for both imaginary and real part of the eigenvalues through their ratio. Since this ratio tells us whether a mode is dominated by transport (far-reaching propagation with little dissipation) or by dissipation (short-range propagation with more dissipation), the ratio filter can emphasize or attenuate modes according to this balance. We now convey this justification in Section 2.4.


We thank again the reviewer for their constructive suggestions and their positive assessment of the analysis. To summarize, following the reviewer’s suggestion, we added an explicit motivation for the two filter families in Section 2.4 and a clearer positioning of our filters against existing spectral GCNs, with the comparison with more recent models following shortly. We hope these changes make clear that the contribution is a new class of graph filters, outside the holomorphic class that existing spectral GCNs share, rather than a variation within it.

AuthorsOpenReview ↗

Follow up on additional experiments

R1.3 Newer baselines and task choice

As announced in our first reply, we report below the comparison with more recent models. Our focus is on spectral GCNs on directed graphs, whose filters can be analyzed as frequency responses of the shift and therefore remain interpretable. Our baselines were primarily chosen to reflect this. Among them, FaberNet in particular reported state-of-the-art results on directed benchmarks. Following the reviewer's suggestion, we broadened the comparison to give a more holistic view of the study, adding the following recent models:

  • FSpecGNN (2026): the architecture is a spectral GNN that incorporates attention heads, resulting in filters that do not commute with the shift operator. As the model is originally applied to undirected graphs, for each dataset we run it on the symmetrized graphs, as in the original paper. We report in particular its Chebyshev variant to facilitate a fair comparison with our approach.
  • DHGNN (2026): the network is a directed message-passing GNN designed for heterophilic graphs and processes the edges and its transpose with two separate learnable encoders.

Both models are tuned under the same protocol as TwinS-GCN. Below, we report the comparison of their performance with respect to our models, indicating accuracy in percentage ± standard deviation

ModelChameleon-FSquirrel-FTexasWisconsinCornell
FSpecGNN (Cheb)44.37 ± 3.8542.55 ± 2.4381.17 ± 5.9884.58 ± 3.9575.77 ± 4.55
DHGNN44.64 ± 3.7942.96 ± 2.6881.26 ± 4.5783.59 ± 4.4371.89 ± 6.30
TwinS-GCN (S)42.33 ± 2.8039.61 ± 1.5584.59 ± 3.4387.12 ± 4.1375.14 ± 3.15
TwinS-GCN (R)44.24 ± 3.2340.15 ± 1.8384.05 ± 3.9187.39 ± 3.9876.76 ± 3.01
TwinS-GCN (C)41.64 ± 3.5940.10 ± 1.5885.41 ± 4.7186.80 ± 3.7975.68 ± 4.01

We also show the overall ranking of all models, including the new baselines, in Table.1 of the revised manuscript. Overall, TwinS-GCN remains competitive with all models and in particular TwinS-GCN (R) has the best mean rank.

We focus on node classification and graph transfer as these tasks most directly benefit from the two properties we claim for our models: long-range reach and non-attenuation. The graph transfer task measures how far the model carries the signal and whether it delivers it without fading or being averaged out. On the other hand, the node classification is the standard benchmark of directed spectral GCNs employed in MagNet and FaberNet. Other benchmarks (e.g, molecular graphs or other long range benchmarks) are less relevant to our setting, as they involve undirected graphs whose shift operators have purely real spectra.


We thank the reviewer for this suggestion, which made our empirical comparison more complete. We hope these additions address the reviewer’s concerns and would be very grateful if the reviewer would consider revisiting their assessment. We remain fully available to clarify any remaining points.


References

(FSpecGNN) Xiaohan Wang, Deyu Bo, Longlong Li, and Kelin Xia. Full-spectrum graph neural networks: Expressive and scalable. In Forty-third International Conference on Machine Learning, 2026.

(DHGNN) Aihu Zhang, Jiaxing Xu, Mengcheng Lan, Shili Xiang, and Yiping Ke. Directed homophily-aware graph neural network. Neural Networks, 206:109807, 2026. ISSN 0893-6080.

We thank the reviewer again for their suggestions, which helped us strengthen the paper. Following the reviewer's comments, and as detailed in our previous replies, we clarified how our filters differ from existing spectral GCNs and why we chose these two filter families, and we added the recent FSpecGNN and DHGNN models (2026), with which TwinS-GCN (R) keeps the best mean rank. As the discussion period closes on November 18 (AoE), we hope these changes address the reviewer's concerns and would be very grateful if they could be taken into account in the final assessment. We would be glad to clarify any remaining point.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised manuscript and tracked changes

We thank all reviewers for their valuable feedback and have uploaded a revised manuscript that with the currently addressed comments. To make the changes easier to follow, the supplementary material includes a version of the manuscript with all modifications highlighted (blue) except for the addition of Figure 1 which is not highlighted. Our individual replies indicate where each change was made. A further revision with the results of the additional experiments will follow.

AuthorsOpenReview ↗

Summary of revisions: sharper theory, matched evaluation, and recent baselines

We thank all reviewers for their constructive feedback, which has substantially strengthened the paper. We are also grateful that the reviewers recognized the operator construction as principled and addressing a real limitation of directed spectral filters (R3), the long-range result of the ratio filter (R3), and the soundness of the analysis and results (R1). The revised manuscript is uploaded, and a version with all changes highlighted is provided in the supplementary material. The main changes are:

Theory, sharpened and extended

  • Scope of the non-attenuation guarantee (R3). Section 3.3.1 now separates what is guaranteed from what is empirical. On any directed graph, no Fourier coefficient is attenuated exponentially. In the vertex domain, the signal norm and the Dirichlet energy are preserved exactly for every normal graph, extending the directed-cycle result of the original submission. New bounds (Appendix C.5) control their deviation on general graphs through a single, computable measure of non-normality, ∥L♯−LT∥F\|\mathbf{L}^\sharp - \mathbf{L}^T\|_F.

  • Spectral conjugate and Drazin inverse (R2). We now explicitly show, right after Definition 2.1, why the nilpotent part of L\mathbf{L} vanishes under conjugate kernel τ\tau, and Section 2.4 motivates the Drazin inverse as the unique operator that makes the ratio filter well defined.

  • Chebyshev reparametrization (R3). It is now justified to improve the conditioning of the filter coefficients, and the word "stable" has been removed.


Positioning and clarity

  • Positioning (R1). Classical spectral GCNs (ChebNet, CayleyNet, ARMA, MagNet, FaberNet) apply holomorphic functions of a single graph operator, which cannot act on dissipation (ℜ(λ)\Re(\lambda)) and transport (ℑ(λ)\Im(\lambda)) separately. The spectral conjugate split solves this restriction while preserving shift invariance on any directed graph, and yields interpretable filters with provable IIR reach on the directed cycle and no exponential attenuation of Fourier coefficients on any graph. The introduction and conclusion now state this positioning explicitly.

  • Motivation of the two filter families (R1). Section 2.4 explains why the sum and ratio kernels are natural choices: the sum kernel controls dissipation and transport separately and the ratio kernel acts on their balance, which no holomorphic kernel can do. Comparison with rational filters (R3). ARMA and Cayley filters are discussed conceptually in Section 3.3.2 and Appendix C.8.

  • Readability (R2). We now provide an outline at the end of the introduction to clarify the general structure of the paper. We restructured Section 2 around a single sequential line of argument: first, Section 2.1 presents spectral oversmoothing as a consequence of real spectra, which motivates isolating the imaginary part of the spectrum, then Section 2.2 presents the Cartesian split as the natural first attempt and shows why it fails. Section 2.3 introduces the spectral conjugate as a novel solution to the problem. Each result is now preceded by a short statement of why it is needed and where it is used, for instance, Corollary 1 is followed by a recall of the goal to find commuting and non-dissipative terms. Figure 1 illustrates the full construction and the architectures.


Experiments

  • Matched evaluation protocol (R3). All directed baselines were re-run with the same splits, validation procedure, search space and trial budget as TwinS-GCN. TwinS-GCN (R) keeps the best mean rank.
  • Recent baselines (R1). We added FSpecGNN (2026) and DHGNN (2026). Of all 13 models of the revised Table 1, TwinS-GCN (R) still has the best mean rank.
  • Cost and scaling (R3). Appendix E reports build time, inference time and memory as graph size grows, and training cost on the benchmarks; the limitations state the small-to-medium scope of the method and propose a route to larger graphs left as outlook.

We sincerely thank the reviewers and the area chair for their time and engagement throughout the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.