acceptodds
Under review as a conference paper at ICLR 2027

Hybrid Diffusion–Transport GNNs on Directed Graphs

Abstract

Directionality is intrinsic to many graph-structured systems, yet the resulting asymmetric operators complicate the spectral foundations of classical graph neural networks. Hermitian Laplacians address this limitation by encoding edge direction through complex-valued phase information while retaining real eigenvalues and a unitary eigenbasis. We reinterpret this construction through a diffusion–transport decomposition, in which the real component captures symmetric diffusion and the imaginary component represents directional transport. Building on this interpretation, we introduce ArcShift, a first-order approximation of the magnetic phase that yields an explicit diffusion–transport propagation operator. We further propose ArcShift-CVNN, a complex-valued extension that jointly transforms the real and imaginary feature representations. Experiments on multiple directed graph benchmarks evaluate the proposed models against the exact magnetic formulation and representative graph-learning baselines in terms of node-classification performance and computational behavior. Overall, ArcShift provides an efficiency- and interpretability-oriented directional propagation operator while maintaining competitive predictive performance.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Machine Learning, Graph Neural Network, Directed Graph, Spectral Processing, Complex-Valued Neural Network

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_7hxGOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a method that combines a decomposition of the magnetic Laplacian into symmetric and directional components, simplifying the complex exponential of the magnetic phase using a first-order Taylor approximation to build ArcShift.

Critical strengths
  • The motivation to decouple the symmetric and directional components of the magnetic Laplacian is interesting, since it allows each component to be modulated independently to understand its contribution to the model.
Critical weaknesses
  • The tested benchmarks are small, and the directional component seems to provide little information as shown in figure 2, where increasing β\beta provides limited benefits on most datasets.
  • The theoretical development is limited and provides little insight into the properties of the proposed model.
Questions for authors
  • The manuscript lacks comparisons with recent directed GNN models [1, 2].
  • Can you evaluate ArcShift on the large datasets from [3]? All datasets used in this paper are small. For instance, experiments with Texas, Cornell, and Wisconsin have very imbalanced classes, and experiments on these datasets can lead to unstable and statistically insignificant results [4].

[1] Sheaves Reloaded: A Directional Awakening, ICLR 2026

[2] Cooperative Sheaf Neural Networks, ICLR 2026

[3] A Simple and Scalable Graph Neural Network for Large Directed Graphs

[4] A critical look at the evaluation of GNNs under heterophily: Are we really making progress?, ICLR 2023

4 comments
AuthorsOpenReview ↗

Response to Reviewer 7hxG

Thank you for your feedback.

W1.

We thank you for this observation. Similar to the magnetic parameter (q) in MagNet, the directional coefficient (\beta) in ArcShift controls the contribution of directional information and is selected based on validation performance.

Increasing (\beta) does not necessarily improve classification accuracy monotonically, just as increasing (q) in MagNet does not necessarily lead to better performance. The optimal contribution of directional information depends on the dataset and its graph structure.

Furthermore, ArcShift-CVNN introduces a learnable coefficient (\gamma), which adaptively adjusts the contribution of directional transport during training, eliminating the need for an explicit directional-parameter grid search.

Therefore, Figure 2 illustrates the sensitivity of ArcShift to directional transport rather than suggesting that stronger directional contributions should consistently improve predictive performance.

W2.

ArcShift builds upon the established magnetic Laplacian framework, which encodes edge direction through complex exponential phase factors while preserving Hermitian structure.

Our theoretical development first expresses the magnetic Laplacian in terms of symmetric diffusion and directional transport components. We then apply a first-order Taylor approximation to the complex exponential phase, yielding a simplified formulation in which directional edge-weight imbalances are represented explicitly through a real antisymmetric transport matrix.

The resulting ArcShift operator remains Hermitian because its diffusion component is real symmetric and its directional transport component is constructed by multiplying a real antisymmetric matrix by the imaginary unit. Furthermore, the first-order transport term depends linearly on directional edge-weight imbalances, rather than through the periodic complex exponential phase.

This simplification provides a mathematically transparent direction-aware propagation operator, while the coefficients (\beta) and (\gamma) enable fixed or adaptive control of directional transport, respectively.

Q1.

We have added DSNN and CSNN to the directed-graph comparison (Table 2) and the mixed directed/undirected comparison (Table 3). These models use the same benchmark datasets and ten predefined splits, with configurations selected by mean validation accuracy.

Q2.

Following the reviewer's suggestion, we have extended our evaluation to arxiv-year and snap-patents. The completed results are shown below. Values are test accuracy (%) reported as mean ± sample standard deviation across five recorded 50/25/25 train/validation/test splits. Models use paired split and training seeds, with configurations selected by mean validation accuracy.

Modelarxiv-yearsnap-patents
ArcShift53.12 ± 0.20Ongoing
ArcShift-CVNN52.31 ± 0.3156.44 ± 0.94

ArcShift evaluation on snap-patents is ongoing. We expect to provide the remaining result within the next one to two days.

For reference, we include selected baseline results from [3] that are relevant to our study. ArcShift and ArcShift-CVNN achieve higher accuracy than most of these baselines on arxiv-year, with ArcShift also outperforming MagNet. On snap-patents, ArcShift-CVNN outperforms all selected baselines with reported results.

Modelarxiv-yearsnap-patents
GCN43.73 ± 0.2239.99 ± 0.35
SGC38.79 ± 0.2235.26 ± 0.04
GPRGNN40.21 ± 0.3332.44 ± 0.25
FSGNN45.99 ± 0.3545.44 ± 0.05
ACMGCN43.30 ± 0.9040.07 ± 0.40
MagNet52.10 ± 0.32—

The published baseline results are taken from [3] for reference. Differences in evaluation protocols may limit direct comparability. A dash indicates that the corresponding result was not reported in [3].

AuthorsOpenReview ↗

Updated Results: Large-Scale Dataset Evaluation [3]

The remaining ArcShift evaluation on snap-patents has been completed. The final results are summarized below.

Modelarxiv-yearsnap-patents
ArcShift53.12 ± 0.2061.26 ± 0.14
ArcShift-CVNN52.31 ± 0.3156.44 ± 0.94

Both ArcShift and ArcShift-CVNN achieve higher reported test accuracy than all selected relevant baselines from [3] on arxiv-year and all those with available results on snap-patents. These findings further support the competitive performance and applicability of the proposed methods on larger directed graphs.

I appreciate the authors' efforts to address my concerns. I will raise my score.

We appreciate your continued consideration of our work.

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

This paper studies directed GNNs. Same as with MagNet, complex-valued phase information is used to indicate the edge direction. A first-order approximation is used to mimic diffusion–transport propagation. Experiments are conducted to compare with baselines before 2024. The quality of this paper is far below ICLR.

Critical strengths
  1. This paper introduces the background on modeling directional information on graphs.
  2. Several datasets have been used.
Critical weaknesses
  1. The quality of this paper is far below the ICLR bar. It looks like the authors do not care at all and did not even try to make the paper 9 pages (large space on page 9 and three large paragraphs are the conclusion).
  2. The novelty is very limited. The formulation basically copies SigMaNet and MagNet.
  3. Poor performance. On 5 out of 7 datasets, the baselines outperform the proposed method. On the remaining 2 datasets, the improvement is marginal. It is really confusing why the paper introduces a method that yields a performance drop.
  4. In terms of interpretability, there is no ground truth to verify it.
  5. In terms of efficiency, the paper should compare with light GNNs, such as SGC and LightGNN.
  6. Missing directed GNN baselines: [1] Graph Learning in 4D: A Quaternion-Valued Laplacian to Enhance Spectral GCNs; [2] MGC: A Complex-Valued Graph Convolutional Network for Directed Graphs
  7. Presentation is poor; e.g., Figure 1 is unnecessary, and there is a lot of space around Figure 1.
Questions for authors

NA

1 comment
AuthorsOpenReview ↗

Response to Reviewer tUjN

Thank you for your feedback.

W1.

We have revised the manuscript to improve its presentation and clarity. In particular, we have expanded the experimental comparisons to include more recent and relevant baselines and provided additional details and supporting analyses in the appendix to strengthen the evaluation of the proposed methods. As shown in our extended experimental results, our proposed models achieve higher average test accuracy than these additional baselines.

  • [1] Sheaves Reloaded: A Directional Awakening (ICLR 2026).
  • [2] Directional Sheaf Hypergraph Networks: Unifying Learning on Directed and Undirected Hypergraphs (ICLR 2026).
  • [3] Cooperative Sheaf Neural Networks (ICLR 2026).
  • [4] Graph Learning in 4D: A Quaternion-Valued Laplacian to Enhance Spectral GCNs.

W2.

We clarify that ArcShift builds upon the magnetic Laplacian framework studied "Graph signal processing for directed graphs based on the hermitian laplacian" by Furutani et al. 2019, while MagNet applies magnetic Laplacian-based operators to graph neural networks. ArcShift is not derived from SigMaNet.

While MagNet employs the exact complex exponential phase to encode edge direction, ArcShift introduces a simple yet expressive first-order formulation that explicitly captures directional edge-weight imbalances through symmetric diffusion and Hermitian directional transport components. This formulation preserves the Hermitian structure while directly representing edge-weight imbalances without the periodicity of the complex exponential phase. The coefficient (\beta) further enables direct control over the contribution of directional transport.

Furthermore, ArcShift-CVNN extends this formulation by learning the transport coefficient (\gamma) jointly with the network parameters, avoiding an explicit directional-parameter grid search.

W3.

We respectfully clarify that the manuscript evaluates 6 datasets, rather than 7. As reported in Table 2, ArcShift-CVNN achieves the highest mean test accuracy on Texas (85.9%) and Wisconsin (86.1%), while ArcShift achieves the highest mean on Cora-ML (80.3%) and the second-highest mean on Telegram (88.6%). Thus, the proposed methods achieve the highest reported mean on three of the 6 datasets.

Importantly, ArcShift is designed as a first-order approximation of magnetic phase-based propagation, rather than a more complex alternative to MagNet. Despite this simplification, ArcShift achieves comparable or even better predictive performance on several datasets. The controlled operator comparison in Table 1 further demonstrates that ArcShift-Core outperforms MagNet-Core in mean test accuracy on four of the six datasets under a matched architecture.

We acknowledge that the proposed methods do not consistently outperform all baselines. Nevertheless, these results demonstrate that the simplified diffusion–transport formulation can preserve competitive predictive performance without requiring the exact magnetic phase.

W4.

Our experiments use real-world directed graph datasets with ground-truth node labels to evaluate predictive performance. However, we clarify that interpretability in our work refers to the mathematical structure of the propagation operator rather than ground-truth explanations of individual predictions.

ArcShift explicitly separates symmetric diffusion, representing graph smoothing, from Hermitian directional transport, capturing directional edge-weight imbalances. The coefficients (\beta) and (\gamma) provide explicit control over the contribution of directional transport in their respective model variants.

Thus, our interpretability claim concerns the transparent mathematical roles of these components, rather than independently validated explanations of individual predictions. We have clarified this distinction in the revised manuscript.

W5.

ArcShift focuses on simplifying magnetic Laplacian-based propagation while explicitly preserving directional information in directed graphs. In contrast, lightweight GNNs such as SGC primarily target computational simplification in general graph-learning settings and do not directly address magnetic directional propagation.

W6.

In response, we have expanded our experimental evaluation to include Graph Learning in 4D: A Quaternion-Valued Laplacian to Enhance Spectral GCNs, along with three additional models published in 2026.

The updated results show that ArcShift generally outperforms the quaternion-valued Laplacian method and the three additional 2026 models.

While we acknowledge MGC as a relevant earlier method, our revised evaluation prioritizes recent developments alongside established magnetic Laplacian-based baselines, particularly MagNet and SigMaNet.

W7.

The manuscript follows the official ICLR template. We have revised the layout to reduce unnecessary whitespace and improve the overall presentation.

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

The authors introduce ArcShift, a directional graph propagation operator derived from a first-order approximation of the magnetic Laplacian. The method decomposes magnetic propagation into a symmetric diffusion component and an antisymmetric directional transport component, allowing the latter to be controlled through an explicit scaling parameter. The authors further introduce ArcShift-CVNN, a complex-valued neural network that jointly transforms real and imaginary representations and learns the transport coefficient during training. The framework is evaluated on six directed node-classification benchmarks against the exact magnetic formulation and several graph neural network baselines.

Critical strengths
  • The authors introduce an intuitive interpretation of magnetic graph propagation by explicitly separating symmetric diffusion and directional transport. Although the decomposition follows directly from Euler's identity, expressing the two components separately makes the contribution of edge direction easier to understand and control
  • The authors provide a relatively comprehensive empirical evaluation, including a controlled comparison with the exact magnetic operator, transport-strength sensitivity analysis, and computational measurements. The proposed models achieve competitive performance across several directed graph benchmarks, while standard ArcShift exhibits lower per-epoch and inference times than the considered magnetic baselines
Critical weaknesses
  • The theoretical novelty of the proposed diffusion–transport decomposition appears limited. The decomposition in Section 3.2 follows directly from Euler's identity, and the resulting separation between symmetric and directional components is already implicit in existing magnetic Laplacian formulations. Moreover, recent directed sheaf approaches, such as [1, 2], have introduced complex-valued directional operators that separate edge orientation from real-valued feature transformations, with [1] recovering the magnetic Laplacian as a special case. The paper does not discuss these closely related approaches or sufficiently clarify the additional theoretical insights provided by its decomposition.
  • The relationship between the proposed ArcShift operator and the first-order approximation of the magnetic Laplacian is not fully established. In Section 3.3, the authors introduce different normalizations for the diffusion and transport components and replace the magnetic phase scaling with an independent coefficient. Consequently, the resulting operator is not simply the first-order approximation of the original magnetic operator. Furthermore, although the construction preserves Hermitian symmetry, it does not necessarily preserve other spectral properties, such as positive semidefiniteness or bounded propagation. The approximation error derived in Theorem 3 concerns the complex exponential rather than the final normalized ArcShift operator, leaving its theoretical relationship to magnetic propagation insufficiently characterized

[1] Mule, Emanuele, et al. "Directional Sheaf Hypergraph Networks: Unifying Learning on Directed and Undirected Hypergraphs." International Conference on Learning Representations, 2026

[2] Fiorini, Stefano, et al. "Sheaves Reloaded: A Direction Awakening." International Conference on Learning Representations, 2026

Questions for authors
  • The proposed diffusion–transport decomposition follows directly from the magnetic Laplacian, while recent directed sheaf formulations [1, 2] already provide complex-valued direction-aware operators. What is the additional theoretical contribution of ArcShift relative to these existing formulations, beyond the first-order approximation of the magnetic phase?
  • Since the final ArcShift operator introduces different normalization schemes and an independent transport coefficient, in what precise sense does it approximate the original magnetic Laplacian? Can the authors provide an approximation-error analysis for the actual normalized propagation operator rather than only for the complex exponential?
  • Figure 2 suggests limited sensitivity to transport strength, and some validation-selected coefficients in Table 9 are close to zero. How much of the predictive performance can be attributed specifically to directional transport? A controlled comparison against an otherwise identical model without the transport component would help clarify its contribution.
  • Both ArcShift and ArcShift-CVNN produce complex-valued representations. Could the authors clarify how standard ArcShift handles the real and imaginary components during activation and classification? Furthermore, could they provide an ablation isolating the effect of complex-valued learnable weights from the other architectural differences, particularly the learned transport coefficient?
  • What explains the reduced per-epoch execution time relative to MagNet, given that the exact magnetic operator can also be precomputed? Could the authors distinguish operator construction costs from propagation costs and clarify whether the computational advantage persists when accounting for hyperparameter selection?
8 comments
AuthorsOpenReview ↗

Response to Reviewer etfM 1/3

Thank you for your feedback.

W1.

Although the diffusion–transport decomposition follows algebraically from Euler's identity, the diffusion and transport components of the exact magnetic formulation remain coupled through the nonlinear phase functions cos⁡(2πqΦ)\cos(2\pi q\Phi) and sin⁡(2πqΦ)\sin(2\pi q\Phi). In particular, the periodicity of the magnetic phase can map distinct directional weight imbalances to identical phase values.

ArcShift addresses this limitation by introducing a transport operator that depends linearly on the directional imbalance Φ=A−A⊤\Phi=A-A^\top, with an independently controllable transport coefficient. This formulation avoids the periodic ambiguity of the magnetic phase in the transport term and allows directional strength to be adjusted without modifying the symmetric diffusion component.

Furthermore, we have expanded our experiments to include the additional baselines from references [1, 2], as you suggested. As noted in our response, our proposed models achieve comparable or higher average test accuracy than these baselines.

W2.

ArcShift is motivated by the first-order approximation of the magnetic phase. Starting from the magnetic Laplacian,

Lq=D−A(s)⊙ei2πqΦ,L_q=D-A^{(s)}\odot e^{i2\pi q\Phi},

the first-order Taylor expansion yields

Lq≈D−A(s)−i2πq(A(s)⊙Φ).L_q\approx D-A^{(s)}-i2\pi q(A^{(s)}\odot\Phi).

This approximation motivates the separation of symmetric diffusion and directional transport.

The final ArcShift propagation operator is defined as

\qquad \beta\ge0.$$ Normalization. The diffusion component $\hat A^n$ employs GCN-style renormalization to support normalized neighborhood aggregation, while the transport component $T^n=A^n\odot\Phi$ retains the original degree normalization to capture directional edge-weight imbalance. These normalization schemes serve distinct roles in the proposed diffusion–transport propagation framework. Independent transport coefficient. In the exact magnetic formulation, the phase parameter $q$ jointly determines the cosine and sine components, coupling symmetric diffusion and directional transport. In contrast, ArcShift introduces an independent coefficient $\beta$ that scales the linear transport term without modifying the symmetric diffusion component. Therefore, $\beta$ is not merely a replacement for $2\pi q$, but a transport-balancing parameter that allows the contribution of directional edge-weight imbalance to be adjusted independently. These modifications extend the first-order approximation into a practical graph propagation operator rather than representing an exact Taylor truncation of the normalized magnetic Laplacian. Spectral properties and empirical validation. ArcShift preserves Hermitian symmetry, ensuring real eigenvalues and a unitary eigenbasis. As explicitly acknowledged in Section 3.5, positive semidefiniteness is not guaranteed, nor is the spectrum necessarily confined to a fixed interval such as $[-1,1]$. Although Theorem 3 characterizes the first-order approximation error of the magnetic phase rather than the final normalized propagation operator, we explicitly compare ArcShift against MagNet, which employs the exact magnetic Laplacian formulation. Our results demonstrate comparable predictive performance across the benchmark datasets, supporting the empirical effectiveness of the proposed approximation despite its relaxed spectral constraints.
AuthorsOpenReview ↗

Response to Reviewer etfM 2/3

Q1.

The primary contribution of ArcShift is not the algebraic diffusion–transport decomposition itself, but the construction of a Hermitian propagation operator that explicitly separates symmetric diffusion from directional transport. In the exact magnetic formulation, both components depend on the same phase parameter through nonlinear trigonometric functions, whose periodicity can map distinct directional weight imbalances to identical phase values.

ArcShift instead employs a transport operator Tn=An⊙ΦT^n=A^n\odot\Phi, retaining linear dependence on directional edge-weight imbalance, together with an independently controllable transport coefficient. Furthermore, ArcShift-CVNN learns the transport coefficient γ\gamma jointly with the network parameters, reducing the need for an external grid search over directional strength.

This learnable parameterization offers a computational advantage in hyperparameter selection. We are also extending our experimental comparisons to the recent directed sheaf-based approaches [1, 2] to evaluate these differences empirically.

Q2.

As clarified in W2, ArcShift is motivated by the first-order approximation of the magnetic phase, while its final propagation operator incorporates GCN-style renormalization and an independently controllable transport coefficient.

To evaluate the effectiveness of this approximation, we conduct a controlled single-layer operator comparison in Section 4.3 (Table 1). MagNet-Core employs the exact magnetic operator, whereas ArcShift-Core uses the proposed first-order diffusion–transport representation. Both models share the same network architecture, node features, data splits, training procedures, and classifiers, isolating the effect of the directional propagation operator.

ArcShift-Core achieves higher mean test accuracy on four of the six datasets: Cornell (63.0% vs. 51.9%), Wisconsin (62.9% vs. 58.8%), Cora-ML (79.9% vs. 73.3%), and CiteSeer (66.0% vs. 58.0%). MagNet-Core performs better on Texas (67.6% vs. 65.9%) and Telegram (51.6% vs. 46.2%).

These controlled results demonstrate that ArcShift can retain useful predictive information from directed graph structure despite replacing the exact magnetic phase with a simplified diffusion–transport operator. Although this comparison does not establish an operator-norm approximation guarantee, it provides empirical evidence supporting the effectiveness of the proposed approximation.

Q3.

We agree that evaluating β=0\beta=0 is important for isolating the contribution of directional transport. In ArcShift,

ArcShiftβn=A^n−iβTn,\mathrm{ArcShift}_\beta^n=\hat A^n-i\beta T^n,

setting β=0\beta=0 removes the directional transport component, yielding diffusion-only propagation.

Figure 2 evaluates the sensitivity of ArcShift to the transport coefficient β\beta over [0,1][0,1] in increments of 0.05, across six datasets and ten predefined splits. Although some datasets favor relatively small positive values of β\beta, the optimal coefficient varies across datasets. Furthermore, validation accuracy is relatively insensitive to the precise transport strength on several datasets, suggesting that the contribution of directional transport is dataset-dependent rather than universal.

The optimal transport coefficient is difficult to determine a priori, similarly to the magnetic phase parameter qq in MagNet, which is typically selected through grid search. Recent directed sheaf-based approaches [1, 2] also involve hyperparameter selection. This motivates our adaptive formulation: ArcShift's linear transport component allows the transport coefficient to be learned jointly with the network parameters. We introduce this learnable coefficient as γ\gamma in ArcShift-CVNN, avoiding a separate grid search over directional strength.

While Figure 2 includes the diffusion-only setting (β=0\beta=0), a direct comparison of its test accuracy against validation-selected positive β\beta under matched experimental settings would further clarify the contribution of directional transport.

AuthorsOpenReview ↗

Response to Reviewer etfM 3/3

Q4.

Activation and classification. Standard ArcShift employs real-valued learnable weights with complex-valued graph propagation, whereas ArcShift-CVNN introduces complex-valued weights that jointly transform the real and imaginary representations.

As described in Section 3.6, ArcShift-CVNN applies a split activation independently to the real and imaginary components:

\qquad H_{\mathrm{im}}=\tanh(Z_{\mathrm{im}}),$$ yielding $H=H_{\mathrm{re}}+iH_{\mathrm{im}}$. For node classification, the hidden representation is normalized row-wise using the L1 norm before the final classification layer: $$\bar H_i=\frac{H_i}{\lVert H_i\rVert_1},$$ with zero rows left unchanged. The normalized representation is then mapped to complex class logits: $$O=\bar H W_{\mathrm{out}}+b_{\mathrm{out}} =O_{\mathrm{re}}+iO_{\mathrm{im}},$$ where $W_{\mathrm{out}}$ and $b_{\mathrm{out}}$ are complex-valued learnable parameters. The real and imaginary logits are aggregated into real-valued class probabilities: $$P=\frac12\left[ \operatorname{Softmax}(O_{\mathrm{re}}) +\operatorname{Softmax}(O_{\mathrm{im}}) \right], \qquad \hat y_i=\arg\max_j P_{ij}.$$ Effect of complex-valued weights. Table 2 shows that ArcShift-CVNN achieves higher mean test accuracy than standard ArcShift on Cornell (81.6% vs. 76.5%), Texas (85.9% vs. 82.7%), Wisconsin (86.1% vs. 83.3%), and CiteSeer (64.6% vs. 64.1%), whereas standard ArcShift performs better on Cora-ML (80.3% vs. 77.5%) and Telegram (88.6% vs. 83.0%). These results suggest that the complex-valued extension can improve predictive performance, particularly on the WebKB datasets. However, ArcShift-CVNN introduces both complex-valued learnable weights and the learned transport coefficient $\gamma$. Therefore, the existing comparison does not independently isolate their contributions. A controlled ablation comparing real- and complex-valued weights under fixed and learned transport coefficients would be necessary to distinguish these effects. **Q5.** We clarify that both MagNet and ArcShift use precomputed graph operators in our experiments. Therefore, the per-epoch training times reported in Table 4 primarily reflect neural network propagation and optimization costs rather than operator construction. These measurements include forward propagation, backward propagation, and parameter updates. Standard ArcShift achieves a lower mean per-epoch execution time than MagNet (16.84 ms vs. 22.95 ms), representing approximately a 26.6% reduction. However, we do not attribute this difference solely to the first-order approximation, as implementation and architectural differences may also contribute. Regarding hyperparameter selection, standard ArcShift evaluates 21 values of $\beta$ over $[0,1]$ in increments of 0.05, whereas MagNet evaluates 6 values of $q$. Across the six datasets, these searches require 336.75 and 121.17 minutes, respectively. Thus, despite its lower per-epoch execution time, standard ArcShift incurs a higher total directional-parameter-selection cost because of its larger search grid. In contrast, ArcShift-CVNN learns the transport coefficient $\gamma$ jointly with the network parameters, requiring only one directional configuration per dataset. Its total measured selection time is 42.96 minutes across the six datasets, compared with 121.17 minutes for MagNet, representing approximately a 64.6% reduction under the reported search protocols. These results distinguish per-epoch training efficiency from directional-parameter-selection cost. While standard ArcShift requires more extensive hyperparameter selection, ArcShift-CVNN reduces this cost through adaptive learning of the transport coefficient. This benefit concerns directional-parameter selection rather than a universal reduction in per-epoch or total training cost.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Manuscript Revisions and Additional Clarifications

Manuscript Revisions

We thank the reviewers for their valuable comments and feedback. We have incorporated their suggestions and revised the manuscript accordingly. The changes are highlighted in yellow to facilitate their identification in the revised manuscript.

1. Expanded experimental comparisons (Tables 2 and 3)

We have added 4 additional graph learning models as baselines:

  • [1] Sheaves Reloaded: A Directional Awakening (ICLR 2026).
  • [2] Directional Sheaf Hypergraph Networks: Unifying Learning on Directed and Undirected Hypergraphs (ICLR 2026).
  • [3] Cooperative Sheaf Neural Networks (ICLR 2026).
  • [4] Graph Learning in 4D: A Quaternion-Valued Laplacian to Enhance Spectral GCNs.

The following table summarizes the average test accuracy across 6 datasets, comparing the four additional baselines with ArcShift and ArcShift-CVNN.

ModelTable 2 (%)Table 3 (%)
DSNN (2026)77.2777.03
DSHN (2026)76.6679.29
CSNN (2026)73.9778.60
Quaternion GCN68.1070.83
ArcShift79.2579.58
ArcShift-CVNN79.7879.67

*DSHN updated on Nov.18

Values represent the unweighted average of mean test accuracy (%) across the six datasets.

Among these 6 models, ArcShift and ArcShift-CVNN achieve slightly higher average test accuracy in both experimental comparisons. These results provide further evidence of the competitive predictive performance of the proposed diffusion–transport formulation.

2. Updated Related Work, Experimental Evaluation, and Appendix

We have updated the Related Work, Experimental Evaluation, and Appendix sections to incorporate four additional studies and their corresponding experimental comparisons. These revisions provide a broader evaluation of ArcShift against recent directional graph-learning approaches, including sheaf-based and quaternion-valued spectral methods.

3. Manuscript layout adjustments

We have made minor layout adjustments to reduce unnecessary whitespace and improve the overall presentation and readability of the manuscript while maintaining the official ICLR formatting requirements.

4. Interpretability and Simplicity of ArcShift

We have reinforced the emphasis on ArcShift's interpretability and simplicity in the manuscript. By applying a first-order approximation to the complex magnetic phase, ArcShift explicitly decomposes Hermitian graph propagation into symmetric diffusion and antisymmetric directional transport components. This formulation provides a clearer interpretation of how graph connectivity and edge direction contribute to feature propagation.

Furthermore, ArcShift simplifies the magnetic propagation operator while preserving its Hermitian structure. The explicit diffusion–transport formulation also allows the directional contribution to be controlled through a single transport coefficient, providing a simple and interpretable alternative to the exact magnetic phase formulation.

Furthermore, ArcShift simplifies the magnetic propagation operator while preserving its Hermitian structure. The explicit diffusion–transport formulation also allows the directional contribution to be controlled through a single transport coefficient, providing a simple and interpretable alternative to the exact magnetic phase formulation.

5. Potential Strength and Applicability of ArcShift

We have further emphasized the potential applicability of ArcShift to weighted directed graphs in which interaction magnitude and directional phase information carry distinct meanings. Unlike conventional magnetic formulations that couple magnitude and phase within complex-valued propagation, ArcShift explicitly separates magnitude-based diffusion and directional transport, allowing their contributions to be controlled independently.

This property may be particularly useful in applications such as transportation, flow, communication, and supply-chain networks, where interaction strength and direction represent different charachteristics of the underlying system. We have highlighted this potential advantage in the revised Conclusion section.

Dear Reviewers,

Now that the authors have responded, please review what they have said and give consideration to whether they have adequately addressed your concerns.

-The AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.