acceptodds
Under review as a conference paper at ICLR 2027

Riemannian Curvature-Transport Spectral Networks for Heterophilous Graph Learning Download PDF

Abstract

Graph neural networks on heterophilous graphs face two coupled challenges: neighborhood aggregation can mix semantically incompatible nodes, and a single latent geometry may fail to capture the coexistence of hierarchical, flat, and cyclic graph structure. Existing methods typically address only one of them: geometric GNNs improve representation fidelity in curved spaces but still rely on smoothing-based aggregation, while heterophily-aware spectral methods improve signal separation yet ignore the geometric mismatch between node representations that live in different tangent spaces. We propose a Curvature-Transport Spectral Network (CTS-Net) that resolves these limitations by combining mixed-curvature representation learning with transport-aware and track-specific propagation. CTS-Net embeds nodes in a mixed-curvature product manifold to model heterogeneous graph geometry, parallel-transports neighbor messages to the receiver tangent space to align cross-node representations before aggregation, routes the aligned messages into smooth and contrastive tracks through a learned orthogonal operator to separate compatible from incompatible interactions, and applies track-specific polynomial diffusion to learn distinct spectral responses for the two types of signals. Our analysis shows that orthogonal routing preserves inter-track separation and that transport avoids additional distortion during aggregation. Across ten heterophilous node-classification benchmarks, CTS-Net achieves the best average performance among the compared baselines while remaining efficient in memory and runtime.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Geometric Graph Neural Network

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_Fmg6OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CTS-Net for learning on heterophilous graphs. The main idea is to address both geometric mismatch and the mixing of incompatible neighbor information during message passing. The model combines mixed-curvature representations, parallel transport for aligning messages across tangent spaces, and a two-track routing mechanism with separate spectral filters. The authors also provide theoretical analysis related to oversmoothing and oversquashing. Experiments on 14 node classification benchmarks show generally strong results, supported by ablation studies and scalability analysis.

Critical strengths
  1. I find the connection between geometric alignment and heterophily particularly interesting. While existing geometric GNNs can represent nodes in curved spaces, they often still rely on smoothing-based aggregation. CTS-Net takes a different approach by explicitly transporting neighbor messages into the receiver's tangent space before aggregation. The use of a shared reference tangent space for subsequent spectral propagation is also a thoughtful design choice, as it keeps messages geometrically consistent across different nodes. I think this gives the paper a clear motivation beyond simply introducing non-Euclidean embeddings.

  2. Another strength is how the different components work together. The routing mechanism uses local structural and geometric information to control how transported messages are distributed between smooth and contrastive tracks. Importantly, the orthogonal routing operator preserves message norms instead of simply suppressing potentially incompatible neighbors. Applying separate polynomial filters to the two tracks then allows the model to learn different spectral responses. I also appreciate that the authors provide theoretical analysis of the norm-preserving properties of transport and routing, as well as the expressive capacity of the independent filters. These details make the framework more convincing from a technical perspective.

  3. Experimental results provide fairly strong support for the proposed design. CTS-Net achieves the best performance among the compared methods on 11 of the 14 benchmarks, with particularly noticeable improvements on datasets such as Texas and Squirrel. The ablation study is also informative. Removing routing leads to the largest average performance drop, while removing transport or replacing the two spectral filters with a single filter also causes clear degradation. These findings help explain why the different components matter. I also find the scalability results encouraging. On Penn94 and ogbn-arxiv, CTS-Net maintains memory usage and runtime comparable to Multi-Track despite the additional geometric operations. Together with the spectral response visualizations and oversmoothing analysis, these experiments provide useful evidence for the proposed framework.

Critical weaknesses
  1. A clearer comparison with existing connection-based GNNs would help establish what is genuinely new in CTS-Net. While CUSP (Grover et al., 2025) and Multi-Track Message Passing (Pei et al., 2024) are discussed, several closely related geometric approaches receive little attention. Sheaf Neural Networks with Connection Laplacians (Barbero et al., 2022, https://proceedings.mlr.press/v196/barbero22a.html) already uses orthogonal maps to align information across neighboring nodes, which is particularly relevant to the proposed parallel transport mechanism. Neural Sheaf Diffusion (Bodnar et al., 2022, https://arxiv.org/abs/2202.04579) also studies geometry-aware propagation under heterophily, although it is only included as an experimental baseline here. More recently, Bundle Neural Network for Message Diffusion on Graphs (Bamberger et al., ICLR 2025, https://openreview.net/forum?id=scI9307PLG) and Gauge-Equivariant Graph Networks via Self-Interference Cancellation (Choi et al., ICML 2026, https://proceedings.mlr.press/v306/choi26d.html) explore related ideas involving geometric transport and graph propagation. I do not consider the combination of transport, routing, and spectral filtering a weakness by itself. However, I would like to see a more direct comparison of how CTS-Net differs from these methods, particularly in its use of mixed-curvature geometry, compatibility-aware routing, and independent spectral propagation.

  2. I would like stronger evidence that geometric alignment itself is responsible for the reported improvements. The ablation results are encouraging, with average Macro-F1 dropping by around 3.2 points without transport and 2.4 points without curvature. However, removing these components also changes the model's representation space and potentially its expressive capacity, making it difficult to isolate the benefit of correcting geometric mismatch. A more direct analysis of how transport changes the compatibility of neighboring representations would help connect the experiments to the paper's central motivation. I also noticed an inconsistency worth clarifying in the curvature settings. Appendix C.3 describes the hyperbolic and spherical curvatures as learnable, yet Table 6 lists the same values of (-1, +1) for every dataset. Are these initialization values or the parameters obtained after training? Reporting the learned curvatures on a few representative datasets would provide more convincing evidence that the mixed-curvature geometry is actually adapting to different graph structures.

  3. Some theoretical conclusions appear broader than what the propositions directly establish. For oversmoothing, the analysis shows that the smooth and contrastive tracks can maintain a certain degree of separation under specific assumptions. However, preserving differences between two tracks does not necessarily prevent node representations within either track from becoming indistinguishable as propagation becomes deeper. Similarly, the oversquashing analysis relies on the norm-preserving properties of parallel transport and orthogonal routing along individual propagation paths. This is useful, but it does not fully address information loss caused by aggregation, path interference, or structural bottlenecks in the graph. I think these results support particular properties of the proposed operators rather than a general guarantee against oversmoothing or oversquashing. A more careful distinction between these theoretical results and the broader empirical claims would make the analysis more convincing.

  4. Figure 1 could do a better job of explaining the main idea of CTS-Net. At the moment, the architecture diagram contains many equations, symbols, and small labels, making it difficult to follow the message flow without repeatedly checking the method section. In particular, the transition from parallel transport to compatibility-aware routing, followed by the shared-reference representation and separate spectral filters, is not immediately clear. I would suggest organizing the figure around these main stages and showing more explicitly where the smooth and contrastive tracks are separated and processed. Moving some of the mathematical details into the main text would also help. Since the interaction between geometry, routing, and spectral propagation is central to the contribution, a clearer overview would make the proposed framework easier to understand.

Questions for authors

Please see Weaknesses 1–3 above. I would particularly appreciate clarification on the differences from related connection-based methods, the learned curvature values, and the scope of the theoretical results.

Reviewer_em2ROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes CTS-Net, which addresses heterophilous graph learning by combining mixed-curvature embeddings, parallel transport, orthogonal routing into two message tracks, and independently learned polynomial filters. Messages are aligned at each receiver, routed, and transported to a common reference tangent space before diffusion. This paper reports improvements across 14 node-classification benchmarks and presents theoretical and empirical analyses of oversmoothing and oversquashing.

Critical strengths
  1. The architecture is well-motivated. Geometric alignment and separation of incompatible neighbor messages address complementary problems. Transporting representations to a common tangent space before ordinary polynomial diffusion is a sensible way to make that stage well-defined.

  2. CTS-Net achieves the highest mean Macro-F1 on 11 of the 14 datasets in Table 1, which shows good improvements over the strongest baselines.

Critical weaknesses
  1. To be honest, the presentation quality of this paper is poor. For example, Figure 1 is a non-vector image and contains a lot of text and equations, making the figure hard to understand. There are also many confusing words. All of these make this paper hard to follow. Moreover, the title ends with "Download PDF". I do not know whether the authors saw their manuscript before submitting.

  2. The geometric operations need clearer explanations. The paper does not fully explain how vectors move between ambient and local tangent coordinates, or how the two tracks are handled during transport. These details are needed to reproduce the method and support its claimed independence from local coordinate choices.

  3. Several statistics, including Questions’ homophily and Tolokers’ edge count, differ from the cited papers. This paper should explain these differences and state whether it removes the known duplicate nodes in Chameleon and Squirrel.

  4. Some datasets lack parameter settings and split details, while Tables 8 and 10 give different results for apparently matching settings. The runtime and memory comparisons also need clearer baseline configurations and preprocessing costs to make the comparisons convincing.

Questions for authors

See weaknesses.

Reviewer_N7sJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses heterophilous graph learning by combining two motivations: neighboring nodes may carry incompatible signals, and heterogeneous graph structure may require different latent geometries. CTS-Net embeds nodes in a product of hyperbolic, Euclidean, and spherical spaces. It parallel-transports neighbor states into the receiver’s tangent space, then uses an orthogonal rotation guided by structural and geometric edge descriptors to route messages into smooth and contrastive tracks. The aggregated states are transported to a common reference tangent space, where independently learned polynomial filters provide different spectral responses before the node update. The paper analyzes transport isometry, inter-track separation, pathwise sensitivity, and polynomial approximation. Experiments on 14 benchmarks report strong overall classification performance, supported by ablations and propagation diagnostics.

Critical strengths
  1. The paper recognizes that improving the geometric representation of a graph does not automatically resolve the mixing of semantically incompatible neighbors. Treating tangent-space alignment and signal routing as separate design problems is a useful perspective. The mixed-curvature representation and explicit transport step provide a concrete way to investigate their interaction within graph message passing.

  2. The closed-form transport operations make the geometric component practical, and parallel transport preserves the Riemannian norm of each individual tangent message. Orthogonal routing similarly preserves the combined message norm in the stated orthonormal coordinates. Independently learned polynomial filters then allow the two tracks to adopt different spectral responses. These are meaningful component-level properties and provide a clear architectural basis for the proposed model.

Critical weaknesses
  1. Equation (5) applies a fixed block rotation in each receiver’s local orthonormal coordinates, but the paper does not define how those frames and the two track subspaces are chosen consistently across nodes. Under a change of local basis, a physical tangent-space operator must transform by conjugation; reusing the same block-rotation matrix generally changes the physical routed message. Even reversing the basis signs in only one track changes the effective rotation if Equation (5) is left unchanged. Transporting the output to the common reference space afterward does not remove this discrepancy. A canonical frame obtained by transporting a fixed reference basis could resolve the issue, but such a construction is not specified. Shared-reference diffusion is well defined once its inputs are fixed; the unresolved issue is the geometric definition of the preceding router and its track decomposition.

  2. Theorem 1 bounds within-node inter-track discrepancy, which can remain positive even when every node has the same representation. Appendix A.2 also assumes retained discrepancy and bounded cancellation through Equations (18)–(20), rather than deriving these properties from orthogonality. More concretely, the Jacobian product in Appendix A.3, Equation (26), omits the learned polynomial filters used by the actual layer. Setting both filters and the skip map to zero makes the layer constant, even with identity layer weights and an activation having a positive derivative at zero, while the displayed lower bound in Equation (9) can remain positive. Thus the isometry argument supports a restricted transport-routing chain, but does not establish the stated sensitivity bound for the complete architecture without additional filter assumptions. The empirical diagnostics are useful, but do not resolve this gap between the analyzed quantities and node-level discriminability or full-network sensitivity.

Questions for authors

see Weaknesses

Reviewer_5pzoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes CTS-Net, a curvature-transport spectral network for heterophilous graph learning. The method combines mixed-curvature representations, parallel transport, orthogonal message routing, and track-specific polynomial spectral filtering. Theoretical analyses and experiments on 14 node-classification benchmarks are conducted to demonstrate the effectiveness of the proposed framework.

Critical strengths
  1. The paper addresses both geometric representation and heterophily-aware message propagation.
  2. The proposed parallel transport mechanism provides a principled way to align representations across tangent spaces.
  3. Experiments cover 14 datasets with different homophily levels and graph scales.
Critical weaknesses
  1. The proposed framework is relatively complex, with several components that increase implementation difficulty.
  2. The theoretical guarantees rely on relatively strong non-degeneracy assumptions.
  3. The improvements on some datasets are marginal, and CTS-Net does not consistently outperform all baselines.
  4. The text in Figure 1 is somewhat blurry, as the figure is not in vector format.
  5. The paragraph starting from Line 62 is too long.
Questions for authors
  1. How sensitive is CTS-Net to the choice of the shared reference point?
  2. How are the smooth and contrastive tracks guaranteed to capture semantically compatible and incompatible information?
  3. What is the additional computational cost of parallel transport compared with conventional message passing?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.