acceptodds
Withdrawn from ICLR 2027

Tango with Drift: A Counteractive Field for Compact Graph Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph representation learning has become a cornerstone of modern machine learning. Yet, inhibiting task-irrelevant and misleading signals on graphs remains highly challenging. Existing methods mainly focus on the static outcome level, \it i.e., by regularizing either graph structures or final representations after detrimental signals have already diffused and become entangled with predictive cues. This not only makes post-hoc removal intrinsically difficult, but also risks collateral damage to task-relevant structures. In this paper, we propose CounterDiff, which casts compact graph learning as a propagation-refinement problem under a graph stochastic partial differential equation (SPDE) view. CounterDiff characterizes the spectral signature of detrimental propagation and enforces coherent counteractions while preserving the predictive flow. We extensively evaluate CounterDiff on diverse graph learning tasks, and the consistent gains highlight the importance of propagation-level intervention for compact yet predictive graph learning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Compact Graph Learning, Graph Representation Learning, Graph Diffusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pQrROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes CounterDiff for compact graph learning under graph stochastic partial differential equations. It introduces a Graph Matérn Counteractive Field that suppresses task-irrelevant signal propagation during message passing. Two complementary losses—a parsimony objective and a sufficiency alignment—regularize the counteractive intervention while preserving predictive information. A Chebyshev polynomial approximation enables efficient computation. Extensive benchmarks show consistent gains over baselines.

Critical strengths
  1. It reframes compact graph learning from static post-hoc compression to intervening during signal propagation, which is conceptually distinct from existing graph information-bottleneck and sparsification methods.
  2. The paper provides rigorous derivations for sufficiency conditions and Chebyshev approximation error bounds for the proposed Matérn counteractive field.
  3. Evaluations span diverse benchmarks and large-graph datasets, with clear comparisons against SOTA models.
Critical weaknesses
  1. Chebyshev polynomial order K trades runtime for accuracy; the paper lacks wall-clock time comparison against competing methods on large graphs.
  2. Performance depends heavily on the weights α\alpha and β\beta for the two losses, without robust automatic tuning strategies provided.
  3. The Graph Matérn Counteractive Field is built upon the undirected graph Laplacian; extension to directed graphs remains unaddressed.
Questions for authors

Please refer to the weaknesses.

Reviewer_dUyeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The author believes that as information is transmitted, unrelated signals in the graph will continue to propagate, and it is too late to compress them after forming the final representation. The author proposed CounterDiff, which adds a Mat é rn random field with graph structure correlation during propagation and constrains unrelated signals with two targets. This paper has been widely validated on 14 graph learning datasets and explained the experimental phenomena using artificially injected high-frequency noise.

Critical strengths
  1. This article places the problem in the propagation process, rather than just dealing with the final output, which has certain research value.
  2. The paper includes multiple datasets such as graph classification, node tasks, and regression, and has also been validated on different backbones, with a relatively complete validation.
  3. The author provides a comprehensive experimental analysis on spectral changes, noise injection, and time and memory overhead.
  4. Table 4 shows that unconstrained random fields actually reduce performance, which is an interesting phenomenon.
Critical weaknesses
  1. Why can zero mean random fields' offset 'irrelevant propagation? This is not explained clearly.
  2. Does compressing the covariance volume really mean deleting irrelevant information? At present, the theory is closer to predicting consistency constraints and is not yet sufficient to support strong explanations for identifying and removing task irrelevant components
  3. What kind of ablation experiments would be lacking if only sufficient constraints were used?
  4. The experimental selection and implementation details are not transparent enough.
  5. The four panels and their explanations in Figure 3 are squeezed together on one line, and the words are noticeably smaller. Tables 1-3 are also crowded on one page.
  6. In terms of writing, there are many repetitions of 'countermeasures, predictive flow, task immediate propagation', but the specific calculation process is not adequately explained.
Questions for authors
  1. Why can zero mean random fields' offset 'irrelevant propagation? Did you really learn about directional cancellation, spectral filtering, or structured noise regularization here?
  2. Does compressing the covariance volume control the representation scale and classification head scale? Can it be proven or experimentally proven that log delta decline is not simply a contraction of amplitude?
  3. If irrelevant signals are located at low frequencies, or if the task signal itself relies on high-frequency structures, is the method still effective?
  4. Is O (K | E |) the complexity of the filtering operation or the complexity of the entire method?
1 comment
Reviewer_dUyeOpenReview ↗

Guidelines on the Use of AI Tools for Manuscript Review

This review process utilized AI to translate and organize the core content of the manuscript. Subsequently, human reviewers conducted a thorough assessment and finalized the review results. Human oversight ensured the overall quality and formatting of the manuscript, and the final conclusions were determined by human judgment rather than by AI.

Reviewer_GJQ1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces CounterDiff, a propagation-level framework for compact graph representation learning. It employs a graph Matérn random field to counteract task-irrelevant propagation, guided by parsimony and predictive sufficiency objectives. The motivation is meaningful, and the results across diverse benchmarks are competitive. Nevertheless, the proposed construction is not adequately connected to the claimed counteractive mechanism, the theoretical derivation contains errors, and the experimental evidence is insufficient to establish the central claims. I currently lean toward rejection.

Critical strengths
  1. Addressing task-irrelevant information during graph propagation, rather than through post-hoc representation compression, offers a meaningful research direction.

  2. The framework accommodates multiple graph architectures, with Table 5 providing preliminary evidence of its transferability across backbones.

  3. The paper supplements predictive results with spectral analyses, noise injection experiments, and ablations to investigate the proposed mechanism.

Critical weaknesses
  1. The motivation for adopting a zero-mean Matérn random field remains insufficient. Topological correlation does not establish task irrelevance, nor does learning a perturbation covariance guarantee the cancellation of detrimental propagation. The paper does not adequately distinguish its proposed counteraction from conventional stochastic regularization or deterministic spectral filtering.

  2. The theoretical analysis seems to contain a verifiable error. For a two-node graph connected by one unweighted edge, setting ν = κ = 1 yields -3 from Eq. (37), whereas Eq. (38) requires -3 ≤ -9. Consequently, the stated Chebyshev error bound is not established by the current proof. Equation (18) also contradicts the sign in Eq. (8). Furthermore, Theorem 1 establishes an information-theoretic identity but does not itself justify the selective removal of task-irrelevant information.

  3. The ablations do not isolate the contribution of the proposed counteractive mechanism. In Table 4, substantial gains emerge only after introducing the sufficiency constraint, while the counteractive field and parsimony objective alone provide no improvement. A full ablation excluding the parsimony objective and comparisons against stochastic perturbations and deterministic filters are needed. Likewise, the high-frequency noise results in Table 6 do not exclude fixed low-pass filtering as an alternative explanation for the observed behavior.

  4. The baseline comparisons are not sufficiently up to date. Although the paper includes some recent methods, many of the compared approaches are relatively old, and the evaluation lacks stronger, more recent state-of-the-art baselines. This weakens the claimed SOTA performance. Moreover, the authors report the best-performing variants across different backbones without clearly specifying the model selection criteria or ensuring comparable parameter counts and hyperparameter search budgets.

  5. The computational analysis does not adequately characterize the full framework. The O(K|E|) complexity in Eq. (14) concerns the spectral operator rather than the complete method, which additionally involves random field generation, ODE integration, predictive consistency, and covariance computations. End-to-end training time, inference latency, peak memory, and performance under matched computational budgets are needed to assess the practical value of the additional complexity.

Questions for authors

See Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.