- 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 -hop limit of degree- 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
-
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.
-
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
-
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.
-
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 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
-
For general graphs, do you explicitly materialize 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?
-
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?
-
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
Follow up on additional experiments
R3.4 Filter implementation and cost
We explicitly compute the operators (, and ), which may be dense, once per graph before training using a single Schur decomposition. The build therefore costs time and 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 ( to regions). Spectral GNNs that use the full spectrum share this cost. Specformer (Bo et al., 2023), for instance, relies on an 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.
| Model | Chameleon-F | Squirrel-F | Texas | Wisconsin | Cornell | Mean rank |
|---|---|---|---|---|---|---|
| MLP | 36.73 ± 4.71 | 36.55 ± 1.82 | 82.97 ± 5.50 | 85.56 ± 4.46 | 76.22 ± 3.91 | 6.80 |
| GCN | 40.89 ± 4.12 | 39.47 ± 1.47 | 52.16 ± 5.16 | 48.92 ± 3.06 | 52.70 ± 5.30 | 10.00 |
| GAT | 39.21 ± 3.08 | 35.62 ± 2.06 | 58.38 ± 6.63 | 49.41 ± 4.09 | 54.32 ± 5.05 | 10.20 |
| MagNet | 40.25 ± 3.84 | 41.27 ± 1.75 | 81.44 ± 4.82 | 83.53 ± 4.22 | 75.05 ± 2.85 | 6.50 |
| Dir-GNN | 43.80 ± 4.43 | 43.48 ± 1.65 | 69.91 ± 6.11 | 71.70 ± 6.44 | 59.46 ± 5.49 | 6.60 |
| FaberNet | 44.51 ± 3.64 | 45.81 ± 2.19 | 83.96 ± 5.36 | 81.76 ± 3.51 | 75.05 ± 4.82 | 3.90 |
| AdvDIFFormer | 42.49 ± 2.94 | 39.83 ± 1.88 | 83.51 ± 4.90 | 79.35 ± 5.21 | 71.53 ± 4.23 | 6.60 |
| ADR-GNN | 45.30 ± 3.81 | 41.34 ± 1.80 | 82.34 ± 4.40 | 85.95 ± 4.12 | 74.14 ± 4.11 | 4.40 |
| TwinS-GCN (S) | 42.33 ± 2.80 | 39.61 ± 1.55 | 84.59 ± 3.43 | 87.12 ± 4.13 | 75.14 ± 3.15 | 4.40 |
| TwinS-GCN (R) | 44.24 ± 3.23 | 40.15 ± 1.83 | 84.05 ± 3.91 | 87.39 ± 3.98 | 76.76 ± 3.01 | 2.60 |
| TwinS-GCN (C) | 41.64 ± 3.59 | 40.10 ± 1.58 | 85.41 ± 4.71 | 86.80 ± 3.79 | 75.68 ± 4.01 | 4.00 |
| *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.
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.
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 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^. 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^ (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, 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.