acceptodds
Under review as a conference paper at ICLR 2027

Representation Alignment of Graph Foundation Models for Cross-Dataset Classification

Abstract

CLIP-based graph foundation models tackle cross-dataset node classification by aligning graph and text embeddings via contrastive learning. However, their practical performance is fundamentally constrained by the distribution shift between source-domain pre-training texts and inference-time classification templates, which causes severe semantic mismatch. To achieve effective cross-domain representation alignment without any fine-tuning, we propose Perturbation Alignment (PA)—a training-free adaptation framework that performs directed perturbations on classification template embeddings guided by subgraph encoding. Specifically, PA applies a single-step perturbation along the composite objective function to enhance the classification margin of the original model, injecting target-domain subgraph information into textual prototypes while preserving pre-trained semantic knowledge. Extensive experiments on a total of seven benchmark datasets demonstrate that PA not only substantially outperforms the vanilla baseline, but also yields particularly significant improvements when source-domain data or template descriptions are missing. Moreover, PA matches or even surpasses lightweight fine-tuning on certain datasets, providing a practical, parameter-efficient, and theoretically principled solution for cross-dataset graph node classification.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph node classification, Representation alignment

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

This paper studies cross-dataset node classification with graph-text foundation models. It proposes Perturbation Alignment (PA), which updates class text prototypes using target subgraph representations while keeping the pretrained encoders frozen. The paper also replaces average pooling with gated pooling. Experiments on seven target datasets show improved average accuracy, especially when class descriptions are unavailable.

Critical strengths

PA is simple and easy to integrate into existing graph-text models because it keeps the pretrained encoders fixed.

The ablation study is informative. PA alone improves the reported average accuracy from 62.01% to 62.94% with full descriptions, and from 53.76% to 57.17% with class names only. The latter setting provides the clearest evidence for the usefulness of PA.

Critical weaknesses

The margin-improvement guarantee is not established. Equation (9) implicitly uses

max_k(a_k - epsilon * b_k)
=
max_k(a_k) - epsilon * max_k(b_k),

which does not hold in general.

There is also ambiguity in the margin definition: the maximum appears to include the correct class, whereas a standard classification margin should use k != y. In addition, the proof does not account for the L2 normalization applied to the perturbed prototypes in Algorithm 1. These issues invalidate the current guarantee and require a corrected analysis or a weaker theoretical claim.

The experimental protocol depends on target-domain validation labels. Appendix A.4 states that epsilon and lambda are selected separately for each target dataset by grid search on the validation set. PA is therefore parameter-update-free, but the reported protocol is not strictly target-label-free. This should be reflected in comparisons with few-shot methods.

There are also reproducibility inconsistencies. Table 6 reports epsilon = 0 for the gated model on Instagram, while Table 3 reports an improvement from 63.04% to 64.20% after adding PA. In addition, Appendix A.4 states that the epsilon search range starts from 0.01, although Table 6 contains several epsilon = 0 entries.

Questions for authors
  1. Can the authors provide a corrected margin analysis that accounts for L2 normalization? Is the maximum in Eqs. (7)-(9) intended to exclude the correct class?

  2. How many labeled target-domain validation samples are used to select epsilon and lambda? Are the few-shot baselines given a comparable target-label budget?

  3. What performance is obtained when the PA hyperparameters are fixed across target datasets or selected only from source-domain data?

  4. Why does Table 6 report epsilon = 0 for the gated Instagram setting while Table 3 reports a clear gain from PA?

  5. Why are epsilon = 0 values reported in Table 6 when Appendix A.4 states that the search range starts at 0.01?

5 comments

Weakness 1: Margin-improvement guarantee not established; Eq. (9) incorrectly separates maximization; ambiguity in margin definition including correct class; L2 normalization ignored; requires corrected analysis.

Response:

We thank the reviewer for identifying these errors. We sincerely apologize for this oversight. We have replaced it with three theorems that establish local objective descent, bound changes in classification margins and error rates, and give a sufficient condition for local risk reduction under normalized updates. The revised statements are reproduced below.

First, we establish that the normalized gradient update decreases the combined objective locally.

Theorem 1 (Objective descent under a normalized gradient update).

Given N≥2N\ge2 unit-norm pretrained class templates Z={zk}k=1N\mathbf Z=\{\mathbf z_k\}_{k=1}^{N}, a unit-norm subgraph encoding h\mathbf h, and the combined objective

L(Z)=λN∑j=1Nh⊤zj+1N(N−1)∑i=1N∑j≠izi⊤zj,\mathcal L(\mathbf Z) = \frac{\lambda}{N}\sum_{j=1}^{N}\mathbf h^\top\mathbf z_j + \frac{1}{N(N-1)} \sum_{i=1}^{N}\sum_{j\ne i}\mathbf z_i^\top\mathbf z_j,

suppose that the current templates are not a stationary point of this objective under the unit-norm constraints. Then a single ordinary gradient-descent step followed by normalization strictly decreases the combined objective for all sufficiently small positive step sizes.

Theorem 1 justifies the normalized PA update as a local descent step for the combined objective.

Second, we characterize how the step size controls PA's effect on classification correctness. Theorem 2 identifies a margin neighborhood containing all samples whose correctness may change, while preserving correct classification for samples with sufficiently large positive margins.

Theorem 2 (Classification stability under a normalized gradient update).

Given N≥2N\ge2 unit-norm class templates Z\mathbf Z and a unit-norm encoding h\mathbf h, apply one gradient-descent step to L\mathcal L in Theorem 1 with λ≥0\lambda\ge0, followed by normalization. For true label y∈{1,…,N}y\in\{1,\ldots,N\}, let m(0):=h⊤zy−max⁡k≠yh⊤zkm(0):=\mathbf h^\top\mathbf z_y-\max_{k\ne y}\mathbf h^\top\mathbf z_k and m(ϵ)m(\epsilon) denote the original and updated margins. For 0<ϵ<N/(λ+2)0<\epsilon<N/(\lambda+2),

∣m(ϵ)−m(0)∣≤4ϵ(λ+2)N.|m(\epsilon)-m(0)|\le\frac{4\epsilon(\lambda+2)}{N}.

Thus any sample with m(0)>4ϵ(λ+2)/Nm(0)>4\epsilon(\lambda+2)/N remains correctly classified. For any target distribution, let R(0)R(0) and R(ϵ)R(\epsilon) be the error rates before and after updating each input separately, with Z\mathbf Z, λ\lambda, ϵ\epsilon, and the tie-breaking rule fixed. Then

\le\Pr\!\left(|m(0)|\le\frac{4\epsilon(\lambda+2)}{N}\right).$$ For fixed $N$ and $\lambda$, increasing $\epsilon$ within the admissible range widens this neighborhood. Its probability mass bounds the possible accuracy change, quantifying the scope for correcting errors and classification stability outside the neighborhood. Finally, we give a sufficient condition for a local reduction in classification risk under normalized perturbations. **Theorem 3 (Local classification-risk reduction under a normalized perturbation).** Given $N\ge2$ unit-norm class templates $\mathbf Z$ and a random unit-norm encoding $\mathbf H$ with true label $Y$, perturb each template by $\epsilon\mathbf v_k(\mathbf H)$ and normalize it, where the directions are bounded, measurable, and independent of $\epsilon$. Let $m(0)$ and $m(\epsilon)$ be the margins defined in Theorem 2, and let $D:=m'(0+)$ denote the right derivative in $\epsilon$. Suppose that $m(0)$ has a density $f_m$ continuous and positive at zero; a version of the conditional distribution of $D$ given $m(0)=t$ converges weakly to the same limit from both sides as $t\to0$; and $m(\epsilon)=m(0)+\epsilon D+o(\epsilon)$ holds uniformly, up to null events, on $\{|m(0)|\le C\epsilon\}$ for every fixed $C>0$ independent of $\epsilon$. Writing $a:=\lim_{t\to0}\mathbb E[D\mid m(0)=t]$, the classification error $R(\epsilon)$ under any fixed tie-breaking rule satisfies $$R(\epsilon)=R(0)-\epsilon f_m(0)a+o(\epsilon).$$ Thus $a>0$ guarantees a strict error reduction for all sufficiently small positive step sizes. Theorem 3 links local accuracy improvement to the mean first-order margin change near the decision boundary. $a>0$ is a sufficient condition for risk reduction. For space considerations, complete proofs of all three theorems appear in **Appendix A.8**. Appendix A.9 checks the boundary condition ($a>0$).

Weakness 2: Experimental protocol depends on target-domain validation labels; ε and λ selected via per-dataset grid search; PA should be reflected in few-shot comparisons.

Response:

We thank the reviewer for raising this concern. We position PA as a training-free cross-dataset generalization method, but relying on target-domain validation labels for hyperparameter selection does limit its zero-shot capability. To evaluate strict zero-shot performance, we add balanced PA, which uses no target labels. Its coefficient is computed from unlabeled data by gradient-norm balancing:

λ=∥∇ZLdeco∥F∥∇ZLcomp∥F+δ≈2.64,\lambda=\frac{\left\|\nabla _ {\mathbf Z}\mathcal L _ {\mathrm{deco}}\right\| _ F}{\left\|\nabla _ {\mathbf Z}\mathcal L _ {\mathrm{comp}}\right\| _ F+\delta}\approx2.64,

where δ>0\delta>0 provides numerical stability. We use the nearest shared candidate λ=2.5\lambda=2.5 and fix ϵ=0.01\epsilon=0.01 for complete descriptions, 0.020.02 for category-only templates, and 0.10.1 when the corresponding source domain is absent. These settings are shared across datasets within each scenario.

Table 1. Accuracy (%) under the fixed, label-free balanced-PA configuration. Desc indicates whether full class descriptions are provided.

MethodDescCora (Academic)History (E-commerce)WikiCS (Wikipedia)Instagram (Social)Avg
None✓67.04 ± 2.3153.13 ± 0.2671.44 ± 0.1663.04 ± 0.3263.66
Compression only✓66.97 ± 1.8852.46 ± 0.2871.74 ± 0.2162.64 ± 0.3863.45
Decorrelation only✓67.04 ± 2.0553.52 ± 0.2771.46 ± 0.2263.30 ± 0.2763.83
Balanced PA✓67.59 ± 2.0553.32 ± 0.2071.27 ± 0.2363.10 ± 0.3163.82
None×66.05 ± 2.2352.80 ± 0.2967.44 ± 0.3143.37 ± 0.6257.42
Compression only×66.73 ± 2.4852.12 ± 0.3167.62 ± 0.3544.06 ± 0.7057.63
Decorrelation only×65.68 ± 2.2750.76 ± 0.3167.21 ± 0.4343.25 ± 0.5256.73
Balanced PA×65.38 ± 1.8855.04 ± 0.2266.77 ± 0.1842.95 ± 0.3357.53

Balanced PA improves the average over no PA from 63.66% to 63.82% with complete descriptions and from 57.42% to 57.53% with class names. The component differences also show why a shared coefficient is a robust default rather than a guaranteed optimum: decorrelation is more useful with complete descriptions, whereas compression is more useful with category names.

Table 2. Accuracy (%) when the corresponding source domain is included or excluded. Balanced PA uses the shared label-free settings; searched PA uses validation labels.

Source domainPACora (Academic)Computers (E-commerce)Instagram (Social)
IncludedNone67.04 ± 2.3163.94 ± 0.5663.04 ± 0.32
ExcludedNone55.35 ± 0.8447.05 ± 0.6847.65 ± 0.99
ExcludedBalanced67.77 ± 0.7752.42 ± 0.1449.24 ± 0.79
ExcludedSearched71.16 ± 0.9752.59 ± 0.3853.95 ± 0.38

When the academic, e-commerce, or social source domain is removed, balanced PA improves Cora, Computers, and Instagram by 12.42, 5.37, and 1.59 percentage points, respectively, without target validation labels. Searched PA is higher on all three because it adapts λ\lambda and ϵ\epsilon to each target setting. The corresponding grouped bar charts are shown in Figure 4(a--b), and the raw values are in Appendix A.7.

Figure 3(b) already compares PA against few-shot methods. Few-shot baselines require gradient-based training to optimize class-specific prompt vectors from labeled examples, whereas PA is entirely training-free—a practical advantage of our approach. Regarding label budgets, num_shot denotes the number of labeled examples per class. PA uses a validation set of 50 samples. At num_shot=10, few-shot methods consume 10×N labeled examples (where N is the number of classes), which substantially exceeds the 50 unlabeled samples used by PA on all datasets except Instagram.

Question 1: Can the authors provide a corrected margin analysis that accounts for L2 normalization? Is the maximum in Eqs. (7)-(9) intended to exclude the correct class?

Response:

Yes. As detailed in our response to Weakness 1, the revised analysis explicitly includes the post-update L2 normalization and defines the signed margin as m=h⊤zy−max⁡k≠yh⊤zkm=\mathbf h^\top\mathbf z_y-\max_{k\ne y}\mathbf h^\top\mathbf z_k. The three revised theorems establish local descent of the normalized objective, a bound on margin and error-rate changes, and a sufficient condition for local risk reduction. Complete proofs are in Appendix A.8.


Question 2: How many labeled target-domain validation samples are used to select ε and λ? Are the few-shot baselines given a comparable target-label budget?

Response:

Figure 3(b) provides the direct few-shot comparison: num_shot denotes the number of labeled examples per class, whereas PA uses a validation set of 50 samples. Thus, at num_shot=10, few-shot methods use 10N10N labeled examples for NN classes, which exceeds PA's 50 unlabeled samples on all datasets except Instagram. Balanced PA uses no target labels and is the strict-zero-shot counterpart.


Question 3: What performance is obtained when PA hyperparameters are fixed across target datasets or selected only from source-domain data?

Response:

This is evaluated by balanced PA in Weakness 2. It fixes λ=2.5\lambda=2.5 across datasets and uses ϵ=0.01\epsilon=0.01, 0.020.02, or 0.10.1 by scenario, without target labels. It improves the average from 63.66% to 63.82% with complete descriptions and from 57.42% to 57.53% with class names (Table 1), and improves Cora, Computers, and Instagram by 12.42, 5.37, and 1.59 points when their corresponding source domains are removed (Table 2). Validation-selected PA is higher because it adapts the two hyperparameters to each target setting.


Question 4: Why does Table 6 report ε = 0 for the gated Instagram setting while Table 3 reports a clear gain from PA?

Response:

We apologize for this inconsistency. It was caused by an error in processing the reported hyperparameter table. The actual setting for gated Instagram used ϵ=0.01\epsilon=0.01, not ϵ=0\epsilon=0, and the corresponding final accuracy has been corrected in the revised manuscript from the erroneous 64.20% to 63.29%.

Question 5: Why are ε = 0 values reported in Table 6 when Appendix A.4 states that the search range starts at 0.01?

Response:

We apologize for the reporting error. The search-range entry and several selected λ\lambda and ϵ\epsilon values in the previous version were affected by the same data-processing issue. In the revised manuscript, we have corrected the search ranges to ϵ∈{0,0.01,0.02,…,0.20}\epsilon \in \{0, 0.01, 0.02, \ldots, 0.20\} and λ∈{0,0.2,0.4,0.6,0.8,1.5,2.5,5,10,100}\lambda \in \{0, 0.2, 0.4, 0.6, 0.8, 1.5, 2.5, 5, 10, 100\}.

We apologize that the initial submission was not as polished as it should have been in terms of experimental depth, theoretical development, and presentation quality. In this revision, we have done our utmost to address these shortcomings through comprehensive refinements across all three dimensions. We are truly sorry for the rough state of the original manuscript, and we deeply appreciate the reviewer's patience and understanding in evaluating the earlier version.

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

This paper proposes Perturbation Alignment (PA) to address the representation mismatch between pre-training texts and inference-time classification templates in CLIP-based graph foundation models. PA uses target subgraph embeddings to guide a single-step perturbation of class text embeddings, aiming to improve class separation without updating pre-trained model parameters. The authors also replace average pooling in GraphCLIP with gated pooling. Experiments on seven cross-dataset node classification benchmarks show that the proposed method improves average accuracy from 62.01% to 63.44%, with larger gains when descriptive class templates are unavailable.

Critical strengths

S1. Simple and effective training-free adaptation with empirical improvements.

The proposed Perturbation Alignment (PA) provides a simple and computationally efficient approach to adapting pre-trained GFMs without updating model parameters. Experiments across seven datasets demonstrate improvements over the GraphCLIP baseline, with particularly notable gains when descriptive class templates are unavailable. The additional experiments under different source-domain training settings further support the practical value of PA for cross-dataset generalization.

Critical weaknesses

W1. Insufficient justification and ablation of the proposed perturbation objective.

The proposed perturbation objective combines subgraph-template similarity reduction and inter-template decorrelation, but the motivation for these specific objectives is not sufficiently justified. In particular, it remains unclear why reducing the similarity between a subgraph and all class templates, including the correct class, should improve representation alignment. Moreover, the ablation study only examines the overall effect of PA without isolating the contributions of its two objectives. Separate ablations and further analysis of their individual roles are needed to substantiate the proposed design.

W2. Dependence on target-domain validation data for hyperparameter selection.

Although PA is presented as a training-free approach for zero-shot cross-dataset classification, its two hyperparameters, λ and ε, are selected through grid search using labeled validation data from each target dataset. Appendix A.4 further shows that the optimal hyperparameters vary substantially across datasets and experimental settings. This raises concerns about the applicability of PA in strictly zero-shot scenarios, where labeled target-domain data may not be available. Experiments with fixed hyperparameters or label-free selection would help address this limitation.

W3. Modest performance improvements and limited evidence of consistent effectiveness.

Although the proposed method improves over GraphCLIP on all seven datasets, the overall gain is relatively modest (62.01% to 63.44%). PA alone improves average accuracy by 0.93 percentage points under the standard template setting, with smaller gains on several datasets. Furthermore, combining PA with gated pooling does not consistently outperform PA alone, as observed on Cora, WikiCS, and Photo in Table 3, raising questions about the additional contribution of gated pooling. While the paper reports means and standard deviations over three test splits, a paired analysis with confidence intervals would help assess the reliability of these improvements.

Questions for authors

Q1. Why does reducing similarity to all class templates, including the correct class, improve representation alignment? Could the authors provide separate ablations of the two perturbation objectives?

Q2. How would PA perform without labeled target-domain validation data? Could the authors evaluate fixed or label-free hyperparameter selection?

Q3. Are the reported improvements statistically significant? Why does combining PA with gated pooling sometimes underperform PA alone?

5 comments

Weakness 1: Insufficient justification and ablation of the proposed perturbation objective; unclear why reducing similarity to all templates (including correct class) should help; lacks separate ablation of the two objectives.

Response:

We thank the reviewer for requesting a clearer justification and separate ablations. We have added a controlled comparison of no PA, compression only, decorrelation only, and balanced PA on four datasets spanning academic, e-commerce, Wikipedia, and social domains, under both complete-description and category-only settings. The results appear as grouped bar charts in Figure 4(b), with the numerical results in Appendix A.7 and reproduced below.

The two objectives address different sources of class confusion. Decorrelation reduces similarity between class templates, encouraging greater separation when their semantics overlap. Semantic compression reduces graph-template similarity to curb overconfidence arising from incorrect semantic interpretations. Since the true class is unknown at inference, compression treats all templates symmetrically. In our approach, reducing aggregate graph-template similarity therefore does not imply reducing every classification margin—when confidence is lowered uniformly across classes while the templates themselves become more separated, the relative ranking of the correct class can still improve, leading to higher classification accuracy.

Table 1. Accuracy (%) with the two objectives separately ablated. Desc indicates whether full class descriptions are provided.

MethodDescCora (Academic)History (E-commerce)WikiCS (Wikipedia)Instagram (Social)Avg
None✓67.04 ± 2.3153.13 ± 0.2671.44 ± 0.1663.04 ± 0.3263.66
Compression only✓66.97 ± 1.8852.46 ± 0.2871.74 ± 0.2162.64 ± 0.3863.45
Decorrelation only✓67.04 ± 2.0553.52 ± 0.2771.46 ± 0.2263.30 ± 0.2763.83
Balanced PA✓67.59 ± 2.0553.32 ± 0.2071.27 ± 0.2363.10 ± 0.3163.82
None×66.05 ± 2.2352.80 ± 0.2967.44 ± 0.3143.37 ± 0.6257.42
Compression only×66.73 ± 2.4852.12 ± 0.3167.62 ± 0.3544.06 ± 0.7057.63
Decorrelation only×65.68 ± 2.2750.76 ± 0.3167.21 ± 0.4343.25 ± 0.5256.73
Balanced PA×65.38 ± 1.8855.04 ± 0.2266.77 ± 0.1842.95 ± 0.3357.53

With complete descriptions, class meanings are better specified, but descriptions can share concepts and wording across classes. Separating these overlapping templates is therefore useful. Decorrelation improves three datasets and leaves Cora unchanged, raising average accuracy from 63.66% to 63.83%. Compression instead lowers the average to 63.45% and hurts three datasets, consistent with useful graph-text alignment being weakened in this setting.

With class names alone, limited context makes class meanings more ambiguous, increasing the risk of confident but incorrect interpretations. Here compression improves Cora, WikiCS, and Instagram, raising the average from 57.42% to 57.63%. Decorrelation hurts all four datasets and lowers the average to 56.73%. These contrasting results support the distinct roles motivating the two objectives.

Balanced PA uses gradient-norm balancing to set a shared coefficient without target labels (details in our response to Weakness 2). It improves average accuracy over no PA in both settings, to 63.82% and 57.53%, respectively. However, balancing gradient magnitudes does not ensure the best trade-off for classification: decorrelation alone performs slightly better on average with complete descriptions, and compression alone with class names. Searched PA selects λ\lambda and ϵ\epsilon using validation data, allowing the relative weighting to match the target setting. It achieves the best observed accuracy among the compared PA configurations on all three source-domain-exclusion benchmarks discussed below. We have clarified the objectives' complementary roles and this distinction between balanced and searched PA in the revised manuscript.

Weakness 2: Dependence on target-domain validation data for hyperparameter selection; λ and ε selected via grid search on labeled target data undermines zero-shot claim; optimal hyperparameters vary substantially across datasets.

Response:

We thank the reviewer for raising this concern. We position our method as an approach to training-free cross-dataset generalization; the validation-selected results use target labels for hyperparameter selection. To evaluate PA under strict zero-shot constraints, we have added a balanced PA strategy that uses no target validation labels, together with the experiments in Figure 4(b) (component ablations) and Figure 4(a) (source-domain exclusion).

Balanced PA sets the relative objective weight by gradient-norm balancing:

λ=∥∇ZLdeco∥F∥∇ZLcomp∥F+δ,\lambda=\frac{\|\nabla _ {\mathbf Z}\mathcal L _ {\mathrm{deco}}\| _ F}{\|\nabla _ {\mathbf Z}\mathcal L _ {\mathrm{comp}}\| _ F+\delta},

where Ldeco\mathcal L_{\mathrm{deco}} and Lcomp\mathcal L_{\mathrm{comp}} are the unweighted decorrelation and compression objectives, and δ>0\delta>0 ensures numerical stability. Computing this ratio on unlabeled data gives λ≈2.64\lambda\approx2.64; we use the nearest candidate, λ=2.5\lambda=2.5, across target datasets. We fix ϵ=0.01\epsilon=0.01 for complete descriptions, 0.020.02 for category-only templates, and 0.10.1 when the corresponding source domain is excluded. Thus each scenario uses a shared configuration across datasets without dataset-specific validation searches.

Component ablations, Figure 4(b). As shown in Table 1 of our response to Weakness 1, balanced PA improves average accuracy from 63.66% to 63.82% with complete descriptions and from 57.42% to 57.53% with class names alone. The separate objectives behave differently: decorrelation is more effective with full descriptions, while compression is more effective with category names. This explains why a shared balance can provide modest average gains without being optimal for every dataset. When the template setting is clearly known, the corresponding single-objective variant also offers a label-free option; balanced PA provides a common configuration when that choice is uncertain.

Source-domain exclusion, Figure 4(a). We remove all training data from the academic, e-commerce, or social source domain and evaluate Cora, Computers, or Instagram, respectively. The numerical results are reproduced below and provided in Appendix A.7.

Table 2. Accuracy (%) with the corresponding source domain included or excluded. Balanced PA uses the shared label-free settings; searched PA uses validation labels.

Source domainPACora (Academic)Computers (E-commerce)Instagram (Social)
IncludedNone67.04 ± 2.3163.94 ± 0.5663.04 ± 0.32
ExcludedNone55.35 ± 0.8447.05 ± 0.6847.65 ± 0.99
ExcludedBalanced67.77 ± 0.7752.42 ± 0.1449.24 ± 0.79
ExcludedSearched71.16 ± 0.9752.59 ± 0.3853.95 ± 0.38

Removing the corresponding source domain substantially reduces accuracy. Despite having no target validation labels, balanced PA improves all three datasets by 12.42, 5.37, and 1.59 percentage points, respectively. Searched PA achieves higher accuracy on all three, showing the additional benefit of adapting the hyperparameters to each target dataset. These results demonstrate that PA can improve transfer under strict zero-shot evaluation, while also quantifying the remaining gap between shared and validation-selected settings. We now distinguish these two protocols explicitly and discuss the limitation of shared hyperparameters in Appendix A.1.

Weakness 3: Modest performance improvements and limited evidence of consistent effectiveness; overall gain 62.01% to 63.44%; PA alone +0.93 pp; gated pooling + PA does not consistently outperform PA alone; lacks statistical significance analysis.

Response:

We thank the reviewer for raising the magnitude and reliability of the improvements. PA's gain is modest with complete descriptions (62.01% to 62.94%, +0.93 points) but becomes larger when information is limited: class names alone yield 53.76% to 57.17% (+3.41 points), including +6.17 points on Instagram and +6.32 points on Photo. With the corresponding source domain excluded, balanced PA improves Cora, Computers, and Instagram by 12.42, 5.37, and 1.59 points, respectively, without target validation labels.

We additionally ran 12 paired evaluations and one-sided paired tt-tests for gated pooling versus the base model and for adding PA to the gated model. Table 3 reports the unadjusted pp-values. The original accuracy tables and figures still use the three-run summaries; due to time constraints, we did not recompute those reported means and standard deviations.

Table 3. One-sided pp-values from paired tt-tests over 12 paired evaluations per dataset. Dashes denote comparisons without a positive mean gain in the original ablation.

ComparisonCoraCiteSeerWikiCSInstagramPhotoComputersHistory
Gated vs. base0.2480.001---0.0010.001
Gated + PA vs. Gated0.0240.0040.0010.0010.001-0.001

Of the four datasets with gated gains, three are significant at the 1% level; all six reported PA gains are significant at the 5% level (five at 1%).

Gated pooling is a small module learned during source-domain training, and its benefits already vary across datasets without PA. Its learned node-importance weights transfer less effectively to target graphs with different neighborhood structures or node-feature distributions. This variability can persist when PA is added, explaining why PA + gated does not consistently outperform PA alone. Nevertheless, with complete descriptions, gated pooling alone improves four of the seven datasets versus three declines, and the combined model achieves higher average accuracy than PA alone.


Question 1: Why does reducing similarity to all class templates, including the correct class, improve representation alignment? Could the authors provide separate ablations of the two perturbation objectives?

Response:

Compression includes all classes because target labels are unavailable. It aims to curb confidence in incorrect semantic interpretations, while decorrelation separates overlapping templates; reducing aggregate similarity does not require equal changes to all class scores. The separate ablations in Table 1 show that decorrelation is more effective with complete descriptions, whereas compression is more effective with class names alone. Please see our response to Weakness 1 and Figure 4(b) for the detailed motivation and analysis.


Question 2: How would PA perform without labeled target-domain validation data? Could the authors evaluate fixed or label-free hyperparameter selection?

Response:

We added balanced PA, using gradient-norm balancing on unlabeled data to set a shared λ=2.5\lambda=2.5 and fixing ϵ\epsilon by scenario, without target validation labels. It improves average accuracy in both template settings (Table 1) and yields gains of 12.42, 5.37, and 1.59 percentage points on Cora, Computers, and Instagram when the corresponding source domain is excluded (Table 2). Please see our response to Weakness 2 and Figure 4(a–b) for the protocol and comparison with searched PA.


Question 3: Are the reported improvements statistically significant? Why does combining PA with gated pooling sometimes underperform PA alone?

Response:

The one-sided paired tt-tests over 12 paired evaluations in Table 3 show significance for three of the four reported gated gains and all six reported PA gains at the 5% level. The original accuracy summaries still use three runs. Gated pooling's source-learned node-importance weights may transfer unevenly across target graphs, so its dataset-dependent benefits can persist when combined with PA. The combination nevertheless improves average accuracy over PA alone. Please see our response to Weakness 3 for the full discussion.

We apologize for the shortcomings in the original manuscript's experiments, theory, and presentation. We have strengthened all three in this revision, including substantial improvements to the aesthetics of all figures and tables, and appreciate the reviewer's constructive feedback.

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

This paper proposes Perturbation Alignment (PA), a training-free adaptation method for cross-dataset node classification with graph foundation models. PA uses target subgraph representations to guide a single-step perturbation of class text embeddings, aiming to reduce cross-domain semantic mismatch. The authors also introduce gated pooling to improve graph representations. Experiments on seven datasets demonstrate improvements over GraphCLIP and other baselines.

Critical strengths
  1. The paper addresses an important challenge in transferring graph foundation models across datasets without task-specific fine-tuning.

  2. PA requires no model parameter updates and introduces negligible inference overhead, making it potentially useful for practical deployment.

  3. Experiments across seven datasets, together with ablation studies on template descriptions and source domains, demonstrate the effectiveness of the proposed approach under different conditions.

Critical weaknesses
  1. The derivation of Theorem 1 appears incorrect, as the maximization operation in Eq. (9) cannot generally be separated as written. Moreover, the proof ignores the L2 normalization applied after perturbation. Therefore, the claimed guarantee of classification margin improvement is not sufficiently justified.

  2. PA requires dataset-specific grid searches over both perturbation strength and objective weight. This dependence weakens its claimed zero-shot applicability and raises concerns about comparison fairness with baselines that do not receive equivalent target-domain validation supervision.

  3. Limited performance gains and novelty. PA alone improves average accuracy by only 0.93 percentage points with complete templates, while the full model gains 1.43 points over GraphCLIP. The improvements are modest on several datasets, and the technical contribution largely consists of a single gradient-based prototype adjustment combined with an existing pooling technique.

Questions for authors

Please refer to weakness.

5 comments

Weakness 1: Derivation of Theorem 1 appears incorrect; maximization cannot be separated as written; L2 normalization is ignored.

Response:

We thank the reviewer for identifying these errors. We agree that the original derivation incorrectly separated the maximization and omitted L2 normalization, so it did not establish the claimed margin-improvement guarantee. We have replaced it with three theorems that establish local objective descent, bound changes in classification margins and error rates, and give a sufficient condition for local risk reduction under normalized updates. The revised statements are reproduced below.

First, we establish that the normalized gradient update decreases the combined objective locally.

Theorem 1 (Objective descent under a normalized gradient update).

Given N≥2N\ge2 unit-norm pretrained class templates Z={zk}k=1N\mathbf Z=\{\mathbf z_k\}_{k=1}^{N}, a unit-norm subgraph encoding h\mathbf h, and the combined objective

L(Z)=λN∑j=1Nh⊤zj+1N(N−1)∑i=1N∑j≠izi⊤zj,\mathcal L(\mathbf Z) = \frac{\lambda}{N}\sum_{j=1}^{N}\mathbf h^\top\mathbf z_j + \frac{1}{N(N-1)} \sum_{i=1}^{N}\sum_{j\ne i}\mathbf z_i^\top\mathbf z_j,

suppose that the current templates are not a stationary point of this objective under the unit-norm constraints. Then a single ordinary gradient-descent step followed by normalization strictly decreases the combined objective for all sufficiently small positive step sizes.

Theorem 1 justifies the normalized PA update as a local descent step for the combined objective.

Second, we characterize how the step size controls PA's effect on classification correctness. Theorem 2 identifies a margin neighborhood containing all samples whose correctness may change, while preserving correct classification for samples with sufficiently large positive margins.

Theorem 2 (Classification stability under a normalized gradient update).

Given N≥2N\ge2 unit-norm class templates Z\mathbf Z and a unit-norm encoding h\mathbf h, apply one gradient-descent step to L\mathcal L in Theorem 1 with λ≥0\lambda\ge0, followed by normalization. For true label y∈{1,…,N}y\in\{1,\ldots,N\}, let m(0):=h⊤zy−max⁡k≠yh⊤zkm(0):=\mathbf h^\top\mathbf z_y-\max_{k\ne y}\mathbf h^\top\mathbf z_k and m(ϵ)m(\epsilon) denote the original and updated margins. For 0<ϵ<N/(λ+2)0<\epsilon<N/(\lambda+2),

∣m(ϵ)−m(0)∣≤4ϵ(λ+2)N.|m(\epsilon)-m(0)|\le\frac{4\epsilon(\lambda+2)}{N}.

Thus any sample with m(0)>4ϵ(λ+2)/Nm(0)>4\epsilon(\lambda+2)/N remains correctly classified. For any target distribution, let R(0)R(0) and R(ϵ)R(\epsilon) be the error rates before and after updating each input separately, with Z\mathbf Z, λ\lambda, ϵ\epsilon, and the tie-breaking rule fixed. Then

\le\Pr\!\left(|m(0)|\le\frac{4\epsilon(\lambda+2)}{N}\right).$$ For fixed $N$ and $\lambda$, increasing $\epsilon$ within the admissible range widens this neighborhood. Its probability mass bounds the possible accuracy change, quantifying the scope for correcting errors and classification stability outside the neighborhood. Finally, we give a sufficient condition for a local reduction in classification risk under normalized perturbations. **Theorem 3 (Local classification-risk reduction under a normalized perturbation).** Given $N\ge2$ unit-norm class templates $\mathbf Z$ and a random unit-norm encoding $\mathbf H$ with true label $Y$, perturb each template by $\epsilon\mathbf v_k(\mathbf H)$ and normalize it, where the directions are bounded, measurable, and independent of $\epsilon$. Let $m(0)$ and $m(\epsilon)$ be the margins defined in Theorem 2, and let $D:=m'(0+)$ denote the right derivative in $\epsilon$. Suppose that $m(0)$ has a density $f_m$ continuous and positive at zero; a version of the conditional distribution of $D$ given $m(0)=t$ converges weakly to the same limit from both sides as $t\to0$; and $m(\epsilon)=m(0)+\epsilon D+o(\epsilon)$ holds uniformly, up to null events, on $\{|m(0)|\le C\epsilon\}$ for every fixed $C>0$ independent of $\epsilon$. Writing $a:=\lim_{t\to0}\mathbb E[D\mid m(0)=t]$, the classification error $R(\epsilon)$ under any fixed tie-breaking rule satisfies $$R(\epsilon)=R(0)-\epsilon f_m(0)a+o(\epsilon).$$ Thus $a>0$ guarantees a strict error reduction for all sufficiently small positive step sizes. Theorem 3 links local accuracy improvement to the mean first-order margin change near the decision boundary. $a>0$ is a sufficient condition for risk reduction. For space considerations, complete proofs of all three theorems appear in **Appendix A.8**. Appendix A.9 checks the boundary condition ($a>0$); its boundary-sample table is reported as **Table 3** in our response to Weakness 3.

Weakness 2: PA requires dataset-specific grid searches; weakens zero-shot applicability and raises fairness concerns.

Response:

We thank the reviewer for raising this concern. We position our method as an approach to training-free cross-dataset generalization, rather than strictly zero-shot inference. Nevertheless, to address the reviewer's concern, we have also evaluated PA under a strict zero-shot protocol without target-domain validation labels.

For this evaluation, we use gradient-norm balancing to set a shared coefficient:

λ=∥∇ZLdeco∥F∥∇ZLcomp∥F+δ,\lambda = \frac{\|\nabla_Z \mathcal{L}_{deco}\|_F}{\|\nabla_Z \mathcal{L} _ {comp}\|_F + \delta},

where δ>0\delta>0 ensures numerical stability. Computing this ratio on unlabeled data gives λ≈2.64\lambda\approx2.64; we use the nearest candidate, λ=2.5\lambda=2.5, across target datasets. We fix ϵ=0.1\epsilon=0.1 when the corresponding source domain is excluded, 0.020.02 for category-only templates, and 0.010.01 for complete templates. These settings are shared across datasets within each scenario, without dataset-specific validation search.

Table 1. Accuracy (%) under fixed settings without target validation labels.

MethodDescCora (Academic)History (E-commerce)WikiCS (Wikipedia)Instagram (Social)Avg
None✓67.04 ± 2.3153.13 ± 0.2671.44 ± 0.1663.04 ± 0.3263.66
Compression only✓66.97 ± 1.8852.46 ± 0.2871.74 ± 0.2162.64 ± 0.3863.45
Decorrelation only✓67.04 ± 2.0553.52 ± 0.2771.46 ± 0.2263.30 ± 0.2763.83
Balanced PA✓67.59 ± 2.0553.32 ± 0.2071.27 ± 0.2363.10 ± 0.3163.82
None×66.05 ± 2.2352.80 ± 0.2967.44 ± 0.3143.37 ± 0.6257.42
Compression only×66.73 ± 2.4852.12 ± 0.3167.62 ± 0.3544.06 ± 0.7057.63
Decorrelation only×65.68 ± 2.2750.76 ± 0.3167.21 ± 0.4343.25 ± 0.5256.73
Balanced PA×65.38 ± 1.8855.04 ± 0.2266.77 ± 0.1842.95 ± 0.3357.53

With complete descriptions, class semantics are better specified, but descriptions of different classes can still overlap in concepts and wording. This motivates decorrelation, which encourages separation among template embeddings to reduce class confusion. The ablation supports this choice: decorrelation raises average accuracy from 63.66% to 63.83%, whereas compression lowers it.

With class names alone, limited context can make labels ambiguous and lead the model to assign high similarity to an incorrect interpretation. Semantic compression is intended to curb such overconfident mismatches. Consistent with this motivation, compression improves three of four datasets and raises average accuracy from 57.42% to 57.63%, whereas decorrelation reduces accuracy on all four.

Balanced PA nevertheless outperforms no PA on average in both settings (63.66% to 63.82% and 57.42% to 57.53%). Thus, under strict zero-shot evaluation, a clearly identified template setting can guide the choice of a PA variant: decorrelation alone for complete descriptions, or compression alone for class names. When that choice is uncertain, balanced PA offers a general default with modest average gains in these experiments.

Table 2. Accuracy (%) under missing source domain.

Source domainPACora (Academic)Computers (E-commerce)Instagram (Social)
IncludedNone67.04 ± 2.3163.94 ± 0.5663.04 ± 0.32
ExcludedNone55.35 ± 0.8447.05 ± 0.6847.65 ± 0.99
ExcludedBalanced67.77 ± 0.7752.42 ± 0.1449.24 ± 0.79
ExcludedSearched71.16 ± 0.9752.59 ± 0.3853.95 ± 0.38

After excluding the academic, e-commerce, or social source domain, balanced PA improves Cora, Computers, and Instagram by 12.42, 5.37, and 1.59 percentage points, respectively, over no PA. Searched PA achieves higher accuracy on all three, illustrating the remaining benefit of validation-based selection. Thus, PA can improve transfer without target validation dataset, while a shared configuration does not match the gains of dataset-specific tuning in these settings. We acknowledge this limitation in Appendix A.1.

We have revised the manuscript to present these results as grouped bar charts in Figure 4(b) (Table 1 above) and Figure 4(a) (Table 2 above), making the differences more directly visible.


Weakness 3: Limited performance gains and novelty; PA alone improves by only 0.93 pp; technical contribution is incremental.

Response:

We thank the reviewer for highlighting the magnitude of the gains. PA is a lightweight inference-time adaptation: it performs one normalized gradient update on the class-template embeddings, introduces no additional training or learnable parameters, and keeps the pretrained model frozen; on Cora, the measured per-batch inference time changes only from 2.05 s to 2.06 s (Appendix A.2). Under complete descriptions, PA alone improves the average accuracy from 62.01% to 62.94% (+0.93 percentage points). The gains become substantially larger under distribution shift: with class names only, PA alone raises the average from 53.76% to 57.17% (+3.41 points), including +6.17 points on Instagram and +6.32 points on Photo. When the corresponding source domain is excluded, the fixed label-free configuration improves Cora, Computers, and Instagram by 12.42, 5.37, and 1.59 points, respectively, while validation-selected PA raises Cora from 55.35% to 71.16% (+15.81 points). These results show that PA is particularly effective when textual descriptions or source-domain coverage is limited.

These gains are also consistent with the local risk-reduction mechanism in Theorem 3. To empirically assess its sufficient condition, we select the samples closest to the original decision boundary within each split and measure the signed margin change Δm. Table 3 reports mean Δm × 10³ over three splits, with the split standard deviation in parentheses:

Table 3. Signed margin changes for samples closest to the original decision boundary.

DatasetNearest 5%Nearest 10%Nearest 20%
Cora (Academic)0.920 ± 0.9001.073 ± 0.3863.005 ± 1.041
History (E-commerce)11.109 ± 0.69210.900 ± 0.55310.655 ± 0.187
WikiCS (Wikipedia)2.133 ± 0.5112.719 ± 0.4703.046 ± 0.469
Instagram (Social)1.243 ± 0.1851.190 ± 0.1651.338 ± 0.229

All twelve estimates are positive, providing empirical support for the (a>0) condition under the tested settings.

PA's technical contribution is an adaptation mechanism for class prototypes in cross-dataset graph classification. It injects target-graph structural information into text prototypes through a joint objective that combines inter-template decorrelation with graph-text semantic compression, followed by a single normalized update. The two objectives address complementary transfer errors—template overlap and incorrect semantic confidence—and their distinct roles are verified by the component ablations. The revised theory explicitly handles L2 normalization and establishes local descent of the composite objective, a classification-stability bound, and a sufficient condition for local risk reduction. Gated pooling is evaluated as a separate architectural component, allowing the contribution of PA to be measured independently. This combination of objective-specific ablations, theoretical analysis, and evaluations under missing descriptions and excluded source domains defines the technical contribution of PA.

We apologize that the initial submission was not as polished as it should have been in terms of experimental depth, theoretical development, and presentation quality. In this revision, we have done our utmost to address these shortcomings through comprehensive refinements across all three dimensions. We are truly sorry for the rough state of the original manuscript, and we deeply appreciate the reviewer's patience and understanding in evaluating the earlier version.

Comments on the paper

1 comment
Area_Chair_p3pUOpenReview ↗

Discussion reminder

Dear Reviewers,thank you for your time and thoughtful evaluation of the Paper. The authors have now submitted their rebuttal addressing the questions and concerns raised in your initial reviews. Please carefully review the authors' responses alongside the other reviews, and participate in the discussion to share your updated thoughts and comments. To help us reach a well-informed meta-review decision, please let us know if the rebuttal adequately resolves your concerns and whether you plan to adjust your final rating. Please submit your post-rebuttal comments and updated evaluation on the review portal. Thank you again for your valuable service and contribution to the review process.

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.