acceptodds
Under review as a conference paper at ICLR 2027

SPR: A Graph Foundation Model for Transferable Graph Cognition through Spectral Patterns and Relational Geometry

Abstract

Recently, Graph Foundation Models (GFMs) have attracted increasing attention for their great potential to learn unified and generalizable knowledge across diverse graphs, thereby supporting a wide range of open-ended graph scenarios. However, unlike natural language and images, graphs are less intuitive for humans to understand. Therefore, despite many initial explorations, a key problem remains unresolved: the transferable cognitive mechanism for graphs. Existing GFMs usually adopt intuitively defined mechanisms to uniformly encode transferable graph patterns, such as handcrafted structural templates (e.g., cycles and trees), fixed propagation operators, and predefined random-walk patterns. However, these mechanisms are often prescriptive and rigid, which may limit their ability to simultaneously model diverse graph patterns. This motivates the exploration of a more flexible, graph-native, and transferable cognitive mechanism. To achieve this, we analyze graphs from the spectral perspective, which provides a natural view of graphs, and propose SPR. SPR decomposes graph information into Chebyshev polynomial bases, facilitating the unified cognition of continuously varying graph spectral patterns. We further model cross-graph relational geometry to enhance the transferability of such graph cognition. Extensive experiments across diverse graphs and downstream scenarios demonstrate that SPR achieves more effective and transferable graph cognition.

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

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

This paper proposes SPR, a graph foundation model combining shared Chebyshev spectral filters with Gromov–Wasserstein prototype geometry for cross-domain transfer. It uses contrastive learning and masked spectral reconstruction for pretraining, and evaluates transfer on node and graph classification, including few-shot settings.

Critical strengths
  1. SPR presents a coherent framework combining shared learnable spectral filters with cross-graph relational geometry to support transfer across heterogeneous graphs.
  2. Under the experimental settings used in the paper, SPR demonstrates strong generalization performance on unseen graphs.
Critical weaknesses
  1. Key details of cross-graph representation alignment remain unclear, including how K-means assignments correspond to fixed prototype indices and how ambiguities in independently computed SVD coordinates are handled. These details are important for assessing the consistency of the shared representation space.
  2. The computational evaluation does not clearly separate geometry construction, barycenter learning, and target alignment costs. Moreover, large-graph experiments replace GW with K-means without a controlled comparison quantifying the resulting accuracy–efficiency trade-off.
Questions for authors

Q1 : On graphs without node attributes, how sensitive is SPR to degree one-hot, constant, random, and structural positional features? Could controlled ablations isolate the contribution of spectral parsing from that of the constructed input features? Q2 : Could you provide controlled comparisons using a randomly initialized encoder, fixed spectral filters, and source-independent prototype geometries, with architecture and preprocessing matched, to isolate the contribution of source-domain learning? Q3 : What does SPR stand for? Please provide its full form at first occurrence, as the acronym is not explicitly defined in the manuscript.

2 comments

We thank the reviewer for the detailed and helpful feedback. We respond to your suggestions point by point as follows.

For W1: We will clarify both alignment mechanisms in §§4.1–4.2 and App. D:

  • Prototype indices: The submitted K-means variant initializes 96 centroids on target nodes and uses their cluster indices directly for top-8 assignment, leaving source-prototype correspondence implicit. Our supplementary variant adds a cluster-level GW coupling that explicitly maps target clusters to the fixed source geometry C⋆C _ \star before K⋆K _ \star propagation.
  • SVD coordinates: Each filtered response Zm,αZ _ {m,\alpha} keeps up to 32 directions in descending singular-value order, fixing signs by making largest-magnitude entries nonnegative (repeated singular values still allow subspace rotations). Even when SVD is computed separately per graph, filter α\alpha retains shared spectral meaning via its Chebyshev coefficients θα,k\theta _ {\alpha,k} prior to SVD, while multi-graph pretraining adapts shared WspecW _ {\mathrm{spec}} to residual basis variation (supported by strict batch size = 1 transfer and Q2 controls).

For W2: Pretraining (Cora, CS, Photo, Chameleon) incurs one-time costs of 4.60 s (diffusion geometry), 14.17 s (GW barycenter C⋆C _ \star), and 77.7 s (25 epochs). Frozen target evaluation (3 splits, 2,000-ep probe) reports GW Dist, Solver, and Total Eval (GW / KM):

Target (nn nodes)GW Dist (s)Solver (s), GW / KMTotal Eval (s), GW / KMAccuracy (%), GW / KM
CiteSeer (3,327)0.1330.562 / 0.0326.46 / 5.9971.89 / 71.05
PubMed (19,717)0.0600.434 / 0.0073.37 / 2.8985.80 / 85.12
Physics (34,493)0.0650.656 / 0.1054.30 / 3.6095.16 / 95.45

Exact GW better preserves relational alignment on CiteSeer and PubMed, whereas K-means cuts solver runtime by an order of magnitude at minimal accuracy cost (slightly improving on denser Physics). This aligns with Prop. 2 and scales to large graphs (ogbn-arxiv, n=169,343n=169,343: 12.5 s total, 1.1 s transport).

For Q1: On IMDB-B (205 matched 1:1:8 splits, val-weighted affine probes; Random Bank = best of 3 norm-matched banks), we compare five structural inputs:

Input on IMDB-BIdentityRandom Bankw/o ProtoSPR
Constant + heat-kernel66.02±2.1066.20±2.1467.24±2.4168.29±2.36
Gaussian (8-ch) + heat-kernel64.12±2.8166.31±2.0567.26±2.8568.35±2.40
One-hot degree + heat-kernel67.84±1.9969.30±1.8269.16±2.2169.61±1.92
LapPE-8 + heat-kernel64.43±2.5666.05±2.1467.03±2.7868.20±2.52
Eight-scale heat-kernel alone64.71±2.3666.46±2.0267.50±2.5468.53±2.40

Full SPR leads every control across all five inputs—including eight-scale heat kernels alone—isolating spectral parsing and prototype transfer beyond structural input choice. Under strict batch size = 1 Table 2 probing, SPR leads all baselines on featureless COLLAB (74.86% vs. 73.78%) and IMDB-B (70.43% vs. 69.60%).

For Q2: We evaluated the contribution of pretraining, prototype geometry, and spectral filtering on two heterophilic graphs (Cornell and Wisconsin) and one homophilic graph (CiteSeer). Results are accuracy (%, mean ± SD).

Model / Control SettingCornellWisconsinCiteSeer
Random Encoder + ER Geometry58.86 ± 2.0267.55 ± 1.9271.84 ± 0.82
Pretrained Encoder + ER Geometry59.00 ± 2.7668.05 ± 2.7570.82 ± 0.73
Prototype Branch Disabled57.19 ± 3.6466.76 ± 2.4071.83 ± 0.74
Identity Filters58.72 ± 1.8868.05 ± 1.9270.70 ± 0.43
Norm-Matched Random Filter Bank59.72 ± 1.8367.05 ± 1.4971.51 ± 0.67
SPR61.24 ± 2.8168.75 ± 2.5072.81 ± 0.73

Replacing the learned prototype geometry with ER geometry or disabling the prototype branch reduces accuracy on all three datasets. On CiteSeer, the pretrained encoder with ER geometry also trails the random encoder with ER geometry, suggesting that the pretrained backbone and gate work best with their learned geometry.

The learned filter bank outperforms both the identity and random banks across the three datasets. Together with the diverse filter responses in Fig. 4(b), these results support the contribution of learned spectral patterns to transfer.

For Q3: SPR stands for Spectral Patterns and Relational Geometry, which we will define explicitly at first use in the Abstract and §1.

We hope these clarifications adequately address your concerns. We are grateful for your feedback.

Thank you for the detailed response and clarifications. I have decided to keep my current score.

Reviewer_3RfJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces SPR, a graph foundation model that composes graph information into Chebyshev polynomial bases, facilitating the unified cognition of continuously varying graph spectral patterns. Afterwrads, SPR models cross-graph relational geometry to enhance the transferability of such graph cognition. The authors report generally strong performance gains, including substantial graph-classification improvements, and provide bounds for geometry reconstruction and the sensitivity of the unnormalized prototype operator.

Critical strengths

The motivation is clear. Figure 1 shows that raw features have different spectral profiles across datasets, which supports using adaptive filters rather than a fixed propagation scheme.

Critical weaknesses
  1. Could the authors provide more detail on the graph-classification setting? With a frozen encoder pretrained only on node-level graphs with 10% of the graphs for training, SPR reports 93.7 on COLLAB, 90.4 on IMDB-B, 91.8 on DD, and 89.2 on PROTEINS, which seems around 10 points above fully supervised GNNs trained on 90% of the data, e.g., GIN and DiffPool, and above self-supervised methods pretrained on the same data, e.g., GraphCL and GraphMAE. If the settings are different, the authors should explicitly disclose the detailed experimental settings and the rationale for adopting a new protocol.
  2. Three reference duplication errors for the papers "Gromov–Wasserstein distances and the metric approach to object matching", OFA, and RiemannGFM.
  3. Representation problems: (i) Figures 3 and 4 are hard to read; (ii) An empty page on page 18.
  4. Some notations are not consistent. For example, λ\lambda denotes Laplacian frequency, while λ∗\lambda_* denotes weights. The definition of DGW\mathcal{D}_{GW} is missing.
  5. The propositions are relatively thin. Specifically, Proposition 1 is a single Jensen step and shows only that the error after lifting is no larger than the GW cost. Proposition 2 is a standard perturbation bound and is not referenced in the main body, which makes it unclear which claim in the method it supports.
Questions for authors

See above.

2 comments

We thank the reviewer for the thoughtful comments. We respond to your concerns as follows.

For W1: We clarify both the evaluation protocol and the resolution of the original ~90% scores:

  • Protocol rationale: §5.2 evaluates the same cross-domain transfer setting across node and graph tasks: encoders are pretrained solely on the four node-level source graphs, kept frozen without target pretraining or finetuning, and evaluated via global mean pooling plus linear probing on 10%/10%/80% stratified splits. This intentionally differs from supervised GNNs (GIN, DiffPool, 90% labels) and target-pretrained SSL (GraphCL, GraphMAE); we will state this explicitly in §5.2 and App. D.
  • Audit and strict independent fix: Auditing our script showed that graphs were originally processed in dataset-order batches sharing a truncated-SVD basis and transport coupling, leaking batch-level transductive cues. We fixed evaluation to strictly independent per-graph encoding (batch size = 1, separate SVD and transport per graph, verified invariant under shuffled graph orderings) under the Table 2 10%/10%/80% linear-probe protocol:
DatasetGCOPEGraphGlueGRACEBGRLGraphMAE2SAMGPTRevised SPR
COLLAB71.41 ± 1.1070.74 ± 1.8972.29 ± 1.1473.78 ± 1.0671.48 ± 1.1371.26 ± 1.2574.86 ± 0.94
IMDB-B66.02 ± 3.2465.55 ± 2.1367.22 ± 2.8067.67 ± 3.9867.10 ± 3.2569.60 ± 1.3770.43 ± 2.31
DD74.00 ± 2.1267.36 ± 1.3564.84 ± 1.3365.15 ± 2.5566.38 ± 2.7465.32 ± 1.8974.59 ± 1.50
PROTEINS72.53 ± 0.6165.90 ± 2.8265.99 ± 4.3063.95 ± 2.8769.11 ± 0.7468.53 ± 1.1373.80 ± 1.52

Under strict per-graph independence with zero target pretraining, SPR still leads across all four benchmarks. Rather than relying on standard message passing—prone to oversmoothing under mean pooling with 10% labels—learned Chebyshev filters preserve discriminative spectral patterns per graph, while frozen K⋆K _ \star and node-wise gating unify local, diffusion, and relational structure across dense social (COLLAB, IMDB-B) and sparse molecular (DD, PROTEINS) graphs. We will update Table 2 and enforce strict per-graph evaluation across §§5.1–5.3.

For W2: Thank you for noticing this issue. We have removed the duplicate bibliography entries for Mémoli (2011), Liu et al. (OFA), and Sun et al. (RiemannGFM).

For W3: Figure 3 shows how accuracy varies with the number of spectral filters and prototypes on CS and Wisconsin. Figure 4 shows the distribution of spectral energy across frequency bands and the responses of the six learned filters. We will enlarge both figures, improve the axis labels, and use distinct line styles and a clearer legend. The whitespace on Page 18 results from the placement of the Appendix B table. We will adjust the table placement and pagination to remove the gap.

For W4: We standardized notation across §4 and App. B: reserving λ∈[0,2]\lambda \in [0,2] for Laplacian frequency and ρ=1−λ∈[−1,1]\rho = 1 - \lambda \in [-1,1] for normalized adjacency eigenvalues (Eq. 5); renaming Eq. 8 source weights from λm\lambda _ m to αm\alpha _ m (αm≥0,∑mαm=1\alpha _ m \ge 0, \sum _ m \alpha _ m = 1) and disambiguating Eq. 1 node distributions from prototype count pp; explicitly defining DGW(Cm,C⋆;Tm)=∑i,j,k,l∣Cm(i,j)−C⋆(k,l)∣2Tm(i,k)Tm(j,l)\mathcal{D} _ {\mathrm{GW}}(C _ m, C _ \star; T _ m) = \sum _ {i,j,k,l} \lvert C _ m(i,j) - C _ \star(k,l)\rvert^2 T _ m(i,k) T _ m(j,l) before Eq. 9; and defining simplices Δn\Delta^n and ∥B∥wm⊗wm2\Vert B\Vert _ {w _ m \otimes w _ m}^2 at first use.

For W5: Both propositions underpin concrete design choices in our two transfer stages. Prop. 1 uses Jensen's inequality to prove DGW(Cm,C⋆;Tm)\mathcal{D} _ {\mathrm{GW}}(C _ m, C _ \star; T _ m) upper-bounds the lifted reconstruction distortion ∥Cm−RmC⋆Rm⊤∥wm⊗wm2\Vert C _ m - R _ m C _ \star R _ m^\top\Vert _ {w _ m \otimes w _ m}^2 (Eq. 9), justifying normalized GW residuals as an alignment confidence indicator (shifting 2.52%–3.23% gate weight from prototypes to local/heat paths under 50% transport perturbation). Prop. 2—now cited after Eq. 13—bounds operator perturbation ∥TK⋆T⊤−T^K⋆T^⊤∥F≤2∥K⋆∥2∥T−T^∥F\Vert T K _ \star T^\top - \hat T K _ \star \hat T^\top\Vert _ F \le 2\Vert K _ \star\Vert _ 2 \Vert T - \hat T\Vert _ F, explaining why sparse K-means (T^\hat T) cuts solver time (0.007–0.105 s vs. 0.434–0.656 s) while matching Sinkhorn GW (TT) accuracy (71.05%/85.12%/95.45% vs. 71.89%/85.80%/95.16% on CiteSeer/PubMed/Physics).

Thanks again for your constructive comments. We hope these clarifications address your concerns.

We sincerely thank you for your time and constructive feedback. As the discussion period will close in approximately one hour, we would like to emphasize that we remain available to address any further questions or concerns. We will be actively monitoring the discussion throughout the final hour and will respond promptly to any additional comments.

We greatly appreciate your consideration and valuable input.

Reviewer_C6kMOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a new perspective for designing graph foundation models based on the self-discovery of transferable graph patterns. Instead of relying on handcrafted subgraph patterns to capture graph invariances and derive transferable units that enable pretrained graph models to generalise to unseen nodes and graphs, they leverage Chebyshev polynomial bases constructed from graph adjacency and degree matrices. These bases provide a continuous relational space for representing graph-specific spectral patterns within a shared embedding space. In parallel, the model learns node-pair relational geometry through diffusion-induced graph distances and a Gromov-Wasserstein transport formulation, enabling the transfer of structural knowledge across graphs. Experimental results on 1-shot and self-supervised node classification, as well as graph-level classification tasks, demonstrate the effectiveness of the proposed approach.

Critical strengths

The SPR model stands out from the existing graph foundation models thanks to its use of the Chebyshev polynomial basis to discover graph invariances. Effective leaning of such basis for each graph and their projection into a shared space, offers a way of discovering structural invariances across diverse graphs, potentially reducing the gap between the universal representation capabilities observed in language and graph model. Furthermore, the Gromov-Wasserstein transport formulation provides an elegant mechanism for learning a transferable prototype geometry that enables cross-graph structural alignment without requiring node identities.

Critical weaknesses

The description of the model architecture lacks clarity. In particular, the main paper does not clearly explain how the different representations are combined or how the prototype geometry enables transfer to unseen graphs.

Section 4.3 refers to augmented graph views, but these views are neither defined beforehand nor referenced to a dedicated description in the appendix, making the training procedure difficult to follow.

The proposed spectral pattern parser relies on node feature representations. This implicit assumption on the availability and quality of node attributes does not hold for many real-world graphs. However, the paper neither discusses this limitation nor evaluates the proposed method in featureless graph settings, making it difficult to assess the generality of the approach.

The proposed relational prototype is learned as an average geometry over the source graphs, yet the paper does not investigate how the framework behaves when the target graph is structurally distant from this prototype. Since the graph-to-prototype transport matrix is central to the method, an analysis of alignment quality and its impact on downstream performance would be valuable. Moreover, the paper provides no evidence that the gate effectively suppresses prototype information for out-of-distribution graphs.

Questions for authors
  • Is there any evidence that the learned spectral filters capture diverse transferable patterns? Do they not collapse to similar values during training?
  • The learned prototype geometry appears to represent an average relational structure derived from the source graphs. How robust is the proposed transfer mechanism when the target graph is structurally distant from the source distribution? In such cases, the Gromov-Wasserstein alignment may yield a weak correspondence between the target graph and the prototype geometry. Can you provide an analysis of transfer quality as a function of the GW distance between the target graph and the prototype, or show whether the gating mechanism effectively suppresses misleading prototype information for out-of-distribution graphs?
  • Also, since the gate does not explicitly incorporate the GW alignment distance to the prototype, it remains unclear whether SPR is robust to strongly out-of-distribution graphs. Insights on this would be welcome.
2 comments

We thank the reviewer for the careful review and constructive suggestions.

For W1: Without target retraining (§§4.1–4.2, Alg. 1), SPR runs three stages: (1) order-3 Chebyshev filters (Mf=6M _ f=6, S=I−LS=I-L) compute Zm,α=Umgα(ρm)Um⊤XmZ _ {m,\alpha}=U _ m g _ \alpha(\rho _ m) U _ m^\top X _ m and rank-32 SVD projects to Xspec∈Rn×128X _ {\mathrm{spec}}\in\mathbb{R}^{n\times 128} (Eqs. 3–6), decoupling θα,k\theta _ {\alpha,k} from raw feature dim; (2) target diffusion aligns to frozen prototypes (C⋆,K⋆)(C _ \star, K _ \star) (p=96p=96) via GW transport TnewT _ {\mathrm{new}} lifting P~new=TnewK⋆Tnew⊤\tilde P _ {\mathrm{new}}=T _ {\mathrm{new}} K _ \star T _ {\mathrm{new}}^\top (Eq. 13, symmetrically normalized to PnewP _ {\mathrm{new}}); (3) over ℓ=1..3\ell=1..3, softmax gates [gA(ℓ),gK(ℓ),gP(ℓ)][g _ A^{(\ell)},g _ K^{(\ell)},g _ P^{(\ell)}] fuse local (HA(ℓ)H _ A^{(\ell)}), heat (HK(ℓ)H _ K^{(\ell)}), and prototype (HP(ℓ)H _ P^{(\ell)}) paths. We will add a §4 overview diagram.

For W2: View 1 (Gm=(Am,Xm)G _ m=(A _ m,X _ m)) masks 25% of Xspec,mX _ {\mathrm{spec},m} nodes for Lspec\mathcal{L} _ {\mathrm{spec}} (Eq. 15); View 2 (G~m=(A~m,Xm)\tilde G _ m=(\tilde A _ m,X _ m)) swaps 5% of edges with non-edges (drop_percent=0.1). Losses balance Lnode\mathcal{L} _ {\mathrm{node}} (1.5), Lspec\mathcal{L} _ {\mathrm{spec}} (0.3), edge BCE (0.2), and DGI (0.4):

ObjectiveChameleonPhotoTexasMean-12
w/o Lnode\mathcal{L} _ {\mathrm{node}}43.68±1.52%92.40±0.36%62.46±1.90%72.28±0.43%
w/o Ldgi\mathcal{L} _ {\mathrm{dgi}}45.02±0.81%92.86±0.39%59.45±6.04%72.84±0.92%
Only Lnode\mathcal{L} _ {\mathrm{node}}45.73±1.43%93.02±0.17%57.26±1.85%71.75±0.42%
Only Lspec\mathcal{L} _ {\mathrm{spec}}43.46±1.20%92.39±0.36%62.33±1.68%72.14±0.42%
SPR46.02±1.45%93.13±0.28%63.56±1.77%73.37±0.55%

All controls trail Full SPR on Mean-12, confirming complementary supervision (§4.3 & App. D).

For W3: When node attributes are absent, SPR replaces XmX _ m with structural inputs X0X _ 0 (degree one-hot, LapPE-8, or heat kernels) before identical filtering, rank-32 SVD, and GW transport: (1) on 205 IMDB-B splits across all five inputs, Full SPR beats Identity, Random, and w/o Proto (eight-scale heat kernels alone: 68.53% vs. 64.71%/66.46%/67.50%; degree + heat: 69.61%), isolating gains beyond input choice; (2) under strict batch-size-1 Table 2 probing, SPR leads all baselines on featureless COLLAB (74.86% vs. 73.78%) and IMDB-B (70.43% vs. 69.60%). Expanded in §4.1 & App. D.

For W4: Replacing 0%–100% of target edges with random non-edges compares Full SPR vs. w/o Proto (paired splits):

Edge ReplacedCornellWisconsinPubMed
0% (Clean)60.55% / 46.48%68.86% / 48.46%87.41% / 83.55%
25% Shift60.14% / 44.55%71.34% / 50.75%85.63% / 74.57%
50% Shift60.55% / 43.31%71.14% / 47.06%84.63% / 68.14%
100% Rewire62.62% / 44.69%72.54% / 51.44%83.97% / 57.70%

Full SPR wins at every level; under heavy rewiring (notably homophilic PubMed), dropping prototypes collapses accuracy whereas Full SPR stays stable via K⋆K _ \star.

For Q1: Filters avoid collapse: centers span λ∈[0.844,1.129]\lambda\in[0.844,1.129] (mean kernel corr -0.189; opposite pairs -0.918, -0.914), separating low/mid/high bands into complementary operators. 100-ep 1-shot ridge probe:

Spectral VariantCornellWisconsinCiteSeer
Identity Filters32.01±7.88%32.53±8.08%49.09±10.45%
Tied Filter Bank32.18±7.78%32.50±8.05%49.55±10.34%
Norm-Matched Random31.01±8.67%30.43±9.22%44.79±9.67%
SPR (Learned Bank)33.18±8.83%33.32±7.54%47.51±9.95%

At 1 label/class, homophilic CiteSeer favors low-pass Identity/Tied priors (too few labels to re-weight bands), whereas heterophilic Cornell/Wisconsin violate low-pass smoothness and gain immediately from diverse bands. With 10% labels to combine all bands (Table 2), SPR beats Identity (+0.70–2.52 pp) and Random (+1.30–1.70 pp) on all three graphs, confirming learned complementary filters drive effective cross-regime transfer.

For Q2&Q3: Extending the base gate (HP(ℓ)H _ P^{(\ell)}) with Prop. 1's node GW residual dGW,id _ {\mathrm{GW},i} gives s~proto,i=sproto,i−βdGW,i\tilde s _ {\mathrm{proto},i}=s _ {\mathrm{proto},i}-\beta d _ {\mathrm{GW},i} (Eq. 9). Under 50% transport perturbation (5 seeds):

DatasetLocal Adj.Heat Diff.Proto PathProto Drop
PubMed47.70%→49.65%31.26%→32.54%21.04%→17.81%-3.23%
Computers42.40%→43.75%42.34%→43.62%15.25%→12.63%-2.63%
Physics45.08%→46.46%40.41%→41.55%14.51%→11.99%-2.52%
CiteSeer44.33%→45.68%40.52%→41.70%15.15%→12.62%-2.53%

Under corrupted transport, the gate down-weights prototypes and shifts weight to local/heat paths; §5 will add GW-vs-transfer curves and this OOD study.

Thanks again for your constructive comments. We hope these clarifications address your concerns.

Thank you for your response. I maintain my overall positive leaning rating.

In response to W4, the authors show that SPR maintains strong performance even under 100% edge rewiring. While this demonstrates robustness to structural perturbations, it also raises the question of how much the model actually relies on graph structure. Since the evaluated datasets (e.g., Cornell, Wisconsin, and PubMed) contain informative node features, it is possible that the model is primarily leveraging node attributes rather than the graph topology.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.