acceptodds
Under review as a conference paper at ICLR 2027

WaveGuard: Coordinating Message Passing and Test-Time Adaptation with Spectral Memory

Abstract

Graph test-time adaptation (GTTA) updates a pretrained graph neural network using an unlabeled target graph without revisiting the source graph. Under structural shift, altered neighborhoods can both degrade node representations and corrupt the signals used for adaptation. Predictive confidence alone does not distinguish these effects, particularly when a model is confidently wrong. We propose WaveGuard, which uses a fixed node-wise intervention score to coordinate message passing and unsupervised updates. A compact class–degree-conditioned memory of source graph-wavelet energy provides a reference for target spectral deviation. Combining this deviation with frozen-model uncertainty yields a score that controls direction-aware soft topology purification and calibrates entropy, anchor, and consistency losses. The score, reweighted graph, and prediction anchors remain fixed during adaptation, while deployment retains only aggregate source statistics. Experiments on five graph benchmarks covering 23 target scenarios demonstrate WaveGuard’s effectiveness under structural, temporal, and mixed shifts. WaveGuard ranks first in six of eight CSBM scenarios, achieves the highest average accuracy on ogbn-arxiv, and outperforms the evaluated baselines in all three PubMed settings and two of three Twitch transfer tasks. On Elliptic, it surpasses the strongest baseline by 6.12 and 6.63 percentage points in Balanced Accuracy and Macro-F1, respectively. Edge-weight analysis and ablation studies demonstrate the complementary contributions of message control and objective calibration.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph test-time adaptation, graph neural networks, structural distribution shift, graph wavelets, spectral memory, source-free adaptation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_piNXOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WaveGuard addresses transductive source-free graph test-time adaptation for node classification. The method computes a fixed node-wise intervention score that combines multiscale graph-wavelet deviation from source statistics with frozen-model predictive entropy, then uses the same score to attenuate directional messages and to weight entropy, anchor, and consistency losses. Experiments cover CSBM, ogbn-arxiv, Elliptic, PubMed, and Twitch, with ablations, edge selectivity analysis, and adaptation dynamics.

Critical strengths
  1. The motivation is grounded in an empirical observation. Figure 1 shows that harmful structural perturbations increase high-frequency energy and spectral centroid, while a homophily-preserving control does not. This supports the claim that multiscale spectral responses can distinguish harmful shift from normal structural heterogeneity, rather than assuming it.

  2. The edge selectivity analysis provides mechanism-level evidence. On PubMed3, clean homophilic edges retain weight 0.956, original heterophilic edges 0.444, and injected heterophilic edges 0.154. This shows the score is selectively attenuating contamination rather than uniformly suppressing heterophily, which is more informative than aggregate accuracy.

Critical weaknesses
  1. The source-reference assumption is not examined at its boundary. The score assumes source spectral statistics remain a meaningful reference for the target. When the target's global topology differs substantially, deviations could be broadly elevated rather than localized, and the score could lose discriminative power. The Twitch DE-to-ES result, where the method trails ActMAD and Matcha-LAME, may reflect this limitation, though the paper attributes it to being target-dependent without analyzing when the source reference ceases to be informative.

  2. The benefit of coupling is not tested against external alternatives. The motivation states that adjusting only the loss or only the structure is insufficient, but the ablations remove internal components rather than comparing against a structure-only or loss-only baseline. A direct comparison would substantiate the coupling claim rather than leaving it as a design rationale.

  3. The incremental value of the source reference and the fixed-score design is not isolated. The paper does not ablate the source reference (e.g., target-only spectral statistics) nor compare against a variant where the score is updated during adaptation. Without these, the gains cannot be attributed specifically to the source reference or to the fixed-score property.

  4. The advantage over adaptation baselines is scenario-dependent. On ogbn-arxiv the margin over SOGA is 0.75 points, and on CSBM4 and CSBM6 WaveGuard ranks second with gaps of 8.2 and 20.2 points. The consistent gains over ERM are clear, but broad superiority over adaptation methods is not observed across all scenarios.

Questions for authors
  1. Can you characterize when the source spectral reference remains informative? For instance, a diagnostic showing how the score distribution behaves as target-source structural distance increases would clarify the boundary of the method's applicability.

  2. Can you compare against a variant that performs structure-only adaptation with standard entropy minimization, and against loss-only adaptation on the original graph, to directly test whether coupling is necessary?

  3. Can you ablate the source reference by using target-only spectral statistics, and compare against a variant where the score is updated during adaptation? This would isolate the contribution of the source reference and the fixed-score design respectively.

  4. On Twitch DE-to-ES, can you analyze why the method trails ActMAD and Matcha-LAME? Understanding this case would help clarify the method's scope.

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

WaveGuard addresses source-free graph test-time adaptation using compact class–degree-conditioned source wavelet statistics. A frozen target model provides spectral deviation and uncertainty scores. The resulting fixed intervention score controls directional message attenuation and weights entropy, frozen-anchor, and consistency objectives. Source-node records are not retained. Experiments cover 23 target scenarios across CSBM, ogbn-arxiv, Elliptic, PubMed, and Twitch. The method reports strong structural- and temporal-shift results, including average Elliptic gains of 6.12 Balanced Accuracy and 6.63 Macro-F1 points over the strongest listed baseline, with more mixed Twitch results (Sections 4–5; Tables 1–3).

Critical strengths
  1. The evaluation spans controlled and naturally occurring shifts, with five-seed comparisons and an explicitly stated target-label-free selection protocol. The Elliptic gains are substantial on both imbalance-sensitive metrics, and the paper candidly reports that Twitch average accuracy is slightly below TSA_T3A. This combination supports a meaningful empirical contribution without requiring a universal adaptation claim (Section 5.1; Tables 1–3; Appendix D).
  2. The two intervention branches receive useful empirical scrutiny. On PubMed3, injected heterophilic edges receive much stronger attenuation than clean homophilic edges, while original heterophilic edges remain partially active. Removing purification or the calibrated objective reduces performance across seven scenarios; degree conditioning and multiscale responses also matter (Sections 5.3–5.4; Figure D2; Table D3). These results make the coupling of message and update control plausible beyond aggregate accuracy alone.
Critical weaknesses
  1. The ablations do not fully isolate the claimed spectral-memory contribution from simpler control scores or independent branch calibration. Table D3 removes score guidance, degree conditioning, or multiscale processing, but does not report the same complete intervention pipeline driven by uncertainty alone, a source feature-distance score, or a global spectral reference with comparable tuning. Nor is shared-score coordination compared with separately calibrated scores for the two branches. Thus, the ablations establish that the implemented score-guided recipe works better than its removals, but not why the proposed conditional spectral memory and shared score are preferable to simpler alternatives (Eqs. (4)–(12); Section 5.4; Table D3).
  2. The handling of confidently wrong predictions remains insufficiently diagnosed despite motivating the method. The class-conditioned distance and threshold are marginalized using frozen class probabilities, so a sharply wrong prediction still selects almost entirely the wrong class reference; soft marginalization mainly helps when uncertainty is appreciable. The anchor reliability also rewards agreement between original and purified frozen predictions, which can be agreement on the same wrong label. Fixing these quantities prevents feedback-driven score changes but can also lock in such mistakes. The paper reports selective edge weights and final accuracy, yet does not show the score and anchor behavior specifically on confidently wrong nodes, leaving this central failure mode unresolved (Introduction; Eqs. (5)–(9); Appendices B.2–B.4).
Questions for authors
  1. What evidence isolates the benefit of conditional spectral memory and a shared intervention score over simpler scores or separately calibrated branches?
  2. How do the fixed score and anchor-reliability mechanism behave on nodes whose frozen predictions are confidently wrong?
1 comment
Reviewer_ZqGmOpenReview ↗

My concerns have been addressed

Thanks for the authors' response. My main concerns have been addressed, and I will raise my score.

Reviewer_8fUDOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

WaveGuard is a source-free graph test-time adaptation method. Offline, it stores class–degree-conditioned statistics of multiscale graph-wavelet energies of the frozen encoder’s hidden features. At test time, a node-wise score ρ_i combines spectral deviation from this memory with predictive entropy; the fixed score then (i) softly down-weights directional messages using cosine-similarity gates and (ii) reweights entropy, anchor-KL and consistency losses. Experiments cover CSBM, ogbn-arxiv, Elliptic, PubMed and Twitch against 12+ TTA/GTTA baselines, with ablations and a sensitivity study.

Critical strengths

Clear motivation: entropy cannot tell confidently-wrong nodes from correct ones, and coupling message control with loss weighting through one fixed score is a reasonable design. Broad evaluation (23 scenarios, 5 seeds), a useful edge-selectivity analysis (Fig. D2) showing injected edges are attenuated far more than natural heterophilic ones, and ablations in which every component contributes. Strong results on Elliptic (+6 BA / +6.6 Macro-F1 over the best baseline) and PubMed.

Critical weaknesses

Baseline quality is questionable. TTReFT and T3R collapse to 13–37% on ogbn-arxiv, 20–40 points below the unadapted model, and T3A reaches 7.5 Macro-F1 on Elliptic3; these look like untuned or broken runs. TENT and TSA_TENT (and T3A and TSA_T3A) report essentially identical numbers on Elliptic and Twitch, which suggests TSA’s adaptation had no effect in your implementation. Baseline sets also differ per dataset: GraphPatcher appears in no table, GOAT and Matcha only on CSBM, Matcha not on ogbn-arxiv. The credibility of the comparison rests on fixing this. Important baselines are missing. The score-calibrated objective is a reliability-weighted entropy loss plus an anchor to the frozen model; EATA (Niu et al., 2022) and SAR (Niu et al., 2023) do exactly this and are standard TTA baselines, yet are absent. A simple similarity-threshold edge reweighting without the spectral score should also be reported, since cosine gating alone may explain the PubMed selectivity result. The spectral component is not isolated. “w/o score guidance” removes the whole score. There is no ablation of spectral deviation vs. entropy alone (β-only), so the paper does not show that the wavelet memory, its central novelty, adds anything beyond frozen-model uncertainty. Figure 1 shows that rewiring changes band energy, not that this signal is predictive of per-node errors beyond entropy.

Questions for authors

Please explain the TTReFT/T3R collapse and the identical TENT vs. TSA_TENT numbers. Were baselines tuned with the same budget? What is the ablation of spectral deviation alone vs. entropy alone in ρ_i? Add EATA/SAR and a plain cosine-threshold reweighting baseline. Which hyperparameters were selected on which domain, and what does “3% adaptation” mean for PubMed?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.