acceptodds
Desk rejected at ICLR 2027

Dual-frequency Decoupling for Exploring Inherent Structural and Topological Features in Homogeneous Graphs

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

While graph transformer methods effectively enlarge the receptive field, the complexity of homogeneous graph data compromises global attention’s ability to capture local structural features. This limitation introduces task-irrelevant long-range topological information, resulting in task representation inaccuracies. To address this challenge, we propose a high- and low-frequency decoupled adaptive constraint method to explicitly capture structural and topological features. First, a dual-frequency decoupling mechanism is designed in the feature capture phase to refine long-range topological information and better distinguish features. Subsequently, residual attention gating augmented by long-range prior knowledge is introduced in the feature integration phase to ensure coherent integration. Finally, multi-scale consistency constraints are proposed in the representation phase to align the integrated representation with the mapping of long-range topological information, ensuring robust training. Our method was benchmarked on six homogeneous datasets, outperforming state-of-the-art methods by 2.15% and 2% on the Cora and Citeseer datasets, respectively.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Node representation learning, Homogeneous Graphs, Dual-frequency decoupling, residual attention gating, Consistency Constraints

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_39YrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a Dual-Frequency Decoupling framework for node classification on homogeneous graphs. The proposed framework consists of three main components: (1) Hierarchical Dual-Frequency Token Encoding (HDTE) separately processes Hop- and Walk-based tokens as local/high-frequency and long-range/low-frequency evidence; (2) Adaptive Hierarchical Global–Local Fusion (AHGF) uses residual gating and adaptively combines the dual-frequency streams; and (3) Multiscale Prior-Interval Constraints (MPIC) imposes metric constraints between the fused representation and the long-range prior while preserving local structural information. Experiments on six homogeneous datasets and three heterogeneous datasets show that the proposed method outperforms the baseline models.

Critical strengths
  1. Strong empirical performance. The method achieves the best results on all six homogeneous graph datasets, and also performs competitive results on the three heterogeneous datasets.
Critical weaknesses
  1. The paper suffers from issues in presentation and logical organization. The paper introduces several concepts and components, while the relationships among them are not clearly established.
  2. Some entries in Table 1 are reported as “unknown values”. This is problematic because Table 1 is the main comparison against existing methods. And it is critical to support the claims. The unusual presentation of the experimental results reduces confidence of the evaluation.
Questions for authors
  1. The organization and logical flow of the paper need great improvement. For example, the title of the paper stresses that the paper focuses on the Homogeneous Graphs, but the authors state that they focus on homogeneous graph data node classification with extensions to heterogeneous graph data in the Method. Additionally, the paper presents Figure 1 to demonstrate that the task representation capability of graph transformer models declines after ablating the long-range represented expression, but the relationship between the ranges and the task representation capability does not seem to be presented.
  2. Table 1 is quite strange, e.g., the unknown values and partially provided second-best results. Besides, some results lack the standard deviation.
  3. The model only works on small datasets, and it should be evaluated on large-scale homogeneous graph datasets.
  4. The symbols are not inconsistent. For example, lisoft\mathcal{l}_i^{soft} is not used in the subsequent process.
  5. In Section 4.4, the authors state that the experiments show the model ability to capture implicit and explicit topological information. So, what is the difference between implicit and explicit topological information? And how is the experiment shown to capture either implicit topological information or explicit topological information?
  6. Hyperparameter analysis is missing.
Reviewer_NN3dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a graph Transformer that separates local structural information from long-range topological information using high- and low-frequency token representations. It then uses adaptive gating to combine these two sources while preserving node-specific features. A multiscale consistency loss further constrains the learned embeddings so that local and global information remain aligned.

The method is evaluated mainly on node classification across six homophilous and three heterophilous graph datasets. It generally outperforms strong GNN and graph Transformer baselines, with especially clear gains on Cora and Citeseer. Ablation studies, token-budget experiments, representation visualizations, and theoretical analysis are used to support the value of the proposed encoding, fusion, and constraint modules.

Critical strengths
  1. The paper identifies a meaningful limitation of graph Transformers: global attention can introduce task-irrelevant long-range information while weakening local structural signals. The proposed high-/low-frequency decomposition gives the architecture an explicit mechanism for preserving these two complementary sources before adaptive fusion.

  2. The method is evaluated across both homophilous and heterophilous node-classification benchmarks, and the ablations show that HDTE, AHGF, and MPIC each contribute to the final performance. The additional walk-sampling, token-budget, and representation analyses strengthen the evidence that the gains are related to the proposed design rather than only to increased model capacity.

Critical weaknesses
  1. The conceptual novelty of the “dual-frequency” formulation is not sufficiently convincing. The separation of local and global information has already been made explicit in GraphGPS, the problem of harmful over-globalization has already been identified and addressed, and adaptive integration of multiple receptive-field scales has subsequently been developed in hierarchical Graph Transformers [1, 2, 3]. Moreover, node-centered multi-hop tokenization has already been introduced by NAGphormer, while hop, walk, and global-context tokens have been jointly modeled by Tokenphormer [4, 5]. Since “high frequency” and “low frequency” are defined here operationally as short-range and long-range information rather than through a graph spectral decomposition, the proposed framework can be viewed largely as a more elaborate routing and fusion mechanism for already established local/global token types. This weakens the case that a sufficiently distinct methodological contribution has been established.

  2. The experimental evidence does not sufficiently establish that the added architectural complexity is necessary or broadly useful. Evaluation is restricted to node classification, while no inductive, graph-level, dynamic-graph, or convincing large-scale efficiency evaluation is provided. In comparison, scalability and computational efficiency have been explicitly evaluated in prior Graph Transformer work, and broader evaluation across numerous datasets and settings has been used to substantiate new architectural designs [1, 6]. This is consequential because several additional mechanisms are introduced simultaneously, separate frequency encoders, cross-frequency attention, multiple gates, prior injection, and additional metric losse without a corresponding parameter-count, training-time, or inference-cost comparison against the strongest baselines. The reported accuracy improvements therefore do not fully establish that the proposed complexity offers a sufficiently favorable or general trade-off.

Presentation Issue: The high-level pipeline can be understood, but substantial effort is required to follow the detailed method because many modules, gates, residuals, priors, and losses are introduced in rapid succession with dense notation. Several presentation issues also reduce clarity, including awkward or incomplete sentences, inconsistent terminology, a “Figure X” placeholder, equation-reference formatting such as “Eq. equation,” and apparently duplicated/mislabeled ablation entries.

References

[1] Recipe for a General, Powerful, Scalable Graph Transformer, NeurIPS'22.

[2] Less is More: On the Over-Globalizing Problem in Graph Transformers, ICML'24.

[3] Unifying and Enhancing Graph Transformers via a Hierarchical Mask Framework, NeurIPS'25.

[4] NAGphormer: A Tokenized Graph Transformer for Node Classification in Large Graphs, ICLR'23.

[5] Tokenphormer: Structure-aware Multi-token Graph Transformer for Node Classification, AAAI'25.

[6] DUALFormer: Dual Graph Transformer, ICLR'25.

Questions for authors
  1. Could it be clarified how the proposed dual-frequency formulation would have differed fundamentally from existing local–global or multi-scale Graph Transformer designs if the high- and low-frequency components were interpreted only as short- and long-range information rather than as a spectral decomposition?

  2. Could it be clarified whether the reported improvements would have remained significant if parameter counts, training time, and inference cost were matched against the strongest Graph Transformer baselines?

  3. Could it be clarified whether the contribution of dual-frequency decoupling itself would have remained significant if the additional adaptive gating and multiscale consistency losses were removed or controlled separately?

  4. Could it be clarified whether similar gains would have been obtained if the method were evaluated beyond transductive node classification, for example under inductive or graph-level settings?

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

This paper studies selective fusion of local structure and long-range context for transductive node classification. Separate Hop/Walk streams interact through attention and hierarchical gates that preserve self-node information and inject a contextualized SGPM graph-level prior. Training combines classification, a rank-dependent triplet loss with Fibonacci-selected negatives, and local cosine alignment. Evaluation covers six predominantly homophilous and three heterophilous graphs; theoretical results concern conditional capacity, bounded fusion, and auxiliary constraints.

Critical strengths
  1. Selective local/global integration. Separate Walk references, local queries, and explicit self-token/prior gates offer a plausible response to irrelevant long-range information overwhelming useful node-specific evidence (§§3.1-3.2).
  2. Promising results across varied settings. Table 1 reports gains over Tokenphormer on all six main datasets, averaging 1.31 percentage points. Table 2 adds heterophilous graphs, while Table 5 and Figure 4 examine module combinations, walk policies, and token budgets. Appendix B specifies the authors' splits and validation-based checkpoint selection.
Critical weaknesses
  1. The specialization mechanism remains insufficiently isolated. Section 3.1 uses “high/low frequency” operationally, rather than as a graph-Fourier decomposition; whether these local/long-range roles explain the gains is therefore central. Table 5's module/walk ablations and Figure 3's Walk-removal comparison do not separate specialization from additional capacity or auxiliary supervision. They establish useful added computation and information, but not why the assigned stream roles are necessary. Table 4 also disables sampling refinement on Cora, Citeseer, and Pubmed, although their Walk branch remains present; these gains cannot validate the context-conditioned refinement. Appendix A establishes potential benefits under complementarity assumptions, not that learned streams realize this regime. Hold the Hop/Walk/SGPM inputs, parameter budget, and losses fixed when comparing specialized interaction against ordinary two-stream fusion, and report refinement on/off results separately. This would identify which mechanism the gains actually support.
  2. Baseline protocols are not sufficiently traceable. Appendix B.1 and Table 1 mix prior-work results with recalculated values. Appendix B.2 specifies stratified 60/20/20 and official Geom-GCN splits, but does not map each baseline entry to its source, split, preprocessing, and rerun status. Consequently, the headline differences cannot yet be confidently attributed to model improvements; the ablation controls do not establish main-table comparability. Annotate copied versus rerun entries and rerun central Tokenphormer/NAGphormer comparisons on identical splits and processed inputs, with comparable tuning budgets and documented graph-prior use. Paired run differences would strengthen the evidence without presuming that any existing baseline result is invalid.
Questions for authors
  1. Under matched inputs, capacity, and losses, what evidence attributes gains specifically to the proposed local/long-range specialization?
  2. Which baseline entries in Tables 1-2 use the exact processed graphs and splits in Appendix B.2?

Comments on the paper

1 comment
AuthorsOpenReview ↗

We sincerely thank the Area Chair and the Reviewers for the careful evaluation of our submission and for the constructive feedback. We addressed all the reviewers' concerns carefully, have expanded our evaluation and clarified several implementation details.

In the following, we highlight the main improvements that happened during the rebuttal period:

  1. Added analysis of high and low frequencies, along with an evaluation of generalization across graph datasets (Section 4.6 and Appendix B). We utilized Fourier transforms to analyze the gain effects of high and low frequencies and examined the integrity of the routing mechanism through spectral analysis. Furthermore, we assessed the model's generalization capability using five datasets: PROTEINS, ENZYMES, MUTAG, PPI, and OGBN-arxiv. The experimental results demonstrate that our proposed method possesses strong generalization capabilities and performs well on graph datasets. These revisions can be found in lines 469–527 of the revised manuscript.
  2. We have standardized the dataset splitting criteria (Appendix B.4). All experimental results have been updated based on these consistent criteria, further validating the effectiveness of our proposed method. The splitting rules for node-level datasets are presented in Table 12, and those for graph-level datasets are in Table 13. (See lines 1193–1237 of the revised manuscript.)
  3. Added an analysis of performance (Appendix B.3). To further investigate the scalability and performance implications of our approach, we conducted additional performance analysis experiments, including comparisons with baseline methods. The results indicate that our method offers distinct advantages. These revisions are located in lines 1148–1183 of the revised manuscript.
  4. Clarified methodological contributions and implementation details (Sections 1 and 3). We have further emphasized the innovations of our work: high-frequency components capture fine-grained structural details and node attributes—though they may overlook long-range dependencies—while low-frequency components capture long-range topological paths and contextual information between nodes. A gating mechanism ensures consistent feature representation. Our approach preserves beneficial long-range signals while suppressing task-irrelevant contextual information. We have also provided further clarification on the specific nature of the high- and low-frequency information. These revisions appear in lines 68–75, 78–83, 198–200, and 223–225 of the revised manuscript.
  5. Expanded ablation study analysis (Appendix B.3). We have extended our ablation study to provide a more comprehensive evaluation of the contributions of our proposed method. These revisions can be found in lines 1034–1058 of the revised manuscript.

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

Related papers

Discussion (0)

Sign in to comment.