acceptodds
Under review as a conference paper at ICLR 2027

SoLIFT: Source–Local Feature Interfaces for Graph Foundation Models

Abstract

Graph foundation models (GFMs) aim to transfer knowledge across graphs, requiring shared encoders to accommodate heterogeneous node attributes. A prominent class of GFM methods maps node attributes to a fixed input width through graph-local spectral transformations. These select feature directions and coordinates independently within each graph. Consequently, recurring source directions may be discarded if they are not locally dominant. We introduce SoLIFT (Source–Local Interfaces For Transfer), a framework that uses source evidence to guide feature selection and coordinate choice. Under an explicit positional feature convention, it combines frozen source subspace evidence with local variation through two variants. SoLIFT-Split reserves common capacity and fills the remaining width with residual PCA, while SoLIFT-Fusion jointly selects directions and orients the selected basis toward a fixed source reference. Our theoretical analysis characterizes how source evidence affects feature selection and shows that coordinate orientation preserves the selected representations' within-graph geometry. We evaluate our framework with seven representative GFM pipelines for transfer from source graphs to unseen target graphs. Our framework improves overall transfer performance over the tested graph-local preprocessing baselines.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph foundation models, cross-graph transfer, node feature preprocessing, spectral methods, subspace learning

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 2 reviews, above 85% of the others.
Reviewer_wV6EOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies feature preprocessing for graph foundation models (GFMs), where graphs may have different feature dimensions and distributions. The authors propose SoLIFT, which uses information from source graphs to guide feature preprocessing on target graphs, rather than applying PCA or SVD to each graph independently. The framework includes two designs SPLIT and FUSION. SPLIT combines source consensus with graph-local residual directions, while FUSION jointly selects source-informed and local feature directions, followed by Procrustes alignment. Experiments across seven GFM pipelines show that FUSION improves the average macro-F1 by 1.97 percentage points over exact SVD.

Critical strengths
  1. The paper addresses a practical problem in graph foundation models: how to handle node features with different dimensions and distributions across graphs. Using information from source graphs rather than preprocessing each graph independently is a reasonable direction.
  2. SoLIFT can be used as a feature preprocessing interface without changing the main architecture of the GFM encoder. This makes it relatively easy to integrate into different models.
  3. The experiments cover seven GFM pipelines, providing a relatively broad evaluation across different architectures. FUSION improves the average macro-F1 from 45.75% with exact SVD to 47.72%, with positive average gains across all seven pipelines.
  4. The paper provides theoretical analysis of the proposed spectral objectives and discusses several limitations, including the assumption of positional feature correspondence and the differences in performance across target datasets.
Critical weaknesses
  1. SoLIFT mainly builds on established techniques, combining source consensus with residual PCA. Although the paper cites relevant prior work, the technical novelty over existing shared/private subspace learning and projector aggregation methods remains unclear. In particular, given that the theoretical results mainly characterize the proposed spectral objectives, does the novelty lie in a new mathematical method, or in combining existing techniques and applying them to GFM feature interfaces?
  2. The paper aims to identify recurring feature directions across different graphs, but node attributes from different graphs may differ in dimensionality, origin, and even semantics. SoLIFT maps them into a common numerical coordinate space without establishing semantic correspondence between the features. Moreover, the paper does not evaluate the impact of this assumption through experiments with independent feature-column permutations across graphs.
  3. Some GFM baselines use different pretraining configurations and modified training protocols. For example, Appendix B.4 shows that GraphGlue uses a shorter, fast training protocol, while other baselines adopt different pretraining configurations. This limits the comparability of performance results across different GFM pipelines. The paper should clarify that the paired comparisons within each pipeline are the main evidence and provide a unified table summarizing training resources and key hyperparameters.
  4. Figure 3 presents t-SNE visualizations of different preprocessing methods on BRIDGE–Cora, but this is a single retrospectively selected case. It cannot be used to directly compare the global representation geometry of different methods quantitatively. The quantitative neighborhood-consistency statistics should be given more emphasis, and the key experimental settings should be included in the figure caption.
Questions for authors
  1. Once FUSION has already selected the feature subspace, why is Procrustes alignment still needed to rotate the basis of the selected subspace? Does this rotation actually improve the transfer performance of GFMs? What is the theoretical justification for this step?
  2. Why are the local subspace and source consensus always assigned equal weights when the relationship between source and target graphs may vary? Could this fixed design lead to negative transfer? The authors explain that both terms have directional support values in the range [0,1], which motivates using equal coefficients. However, having the same numerical range does not imply the same predictive reliability. A more detailed explanation of this design choice is needed.
  3. Although the paper provides relatively extensive performance comparisons across seven GFM pipelines, the existing experiments mainly focus on the final classification performance of different feature preprocessing methods. A more comprehensive ablation study of coordinate orientation is needed to evaluate its actual contribution across different GFM architectures and whether its effectiveness is consistent. In addition, a sensitivity analysis of the source–local fusion weights would help assess the reasonableness of the fixed equal-weight design in FUSION and its impact on model performance.
4 comments
AuthorsOpenReview ↗

Thanks for your comment!

We appreciate your careful assessment of SoLIFT's practical relevance and your detailed questions about its contribution and component effects. We have added controlled studies and clarified how the theory, cross-pipeline comparisons and representation analysis support our conclusions.

W1. Contribution beyond shared/private subspace learning

The contribution is a source–local GFM interface, its deployment protocol and analysis. SoLIFT fits spectral source evidence before backbone training, then combines it with each unseen graph's local variation within a fixed input width. Split explicitly reserves common capacity and fills the remainder with residual local directions; Fusion jointly allocates the full width. Projector averaging, residual PCA and Procrustes are established ingredients; we now state this explicitly.

In Knowledge Transfer across Multiple PCA Studies (Li et al., Algorithm 1), the target participates in the shared-space estimate alongside informative sources, followed by target-private residual PCA. Personalized PCA (Shi and Al Kontar, JMLR 2024) jointly optimizes shared and local reconstruction components for participating clients. SoLIFT estimates its common state from sources alone and reuses it across unseen targets without refitting the source state. This operational distinction makes simpler source-informed estimators relevant controls.

Our analysis characterizes the source–local tradeoff. At equal weights and matched rank, Fusion's gain in consensus overlap is at least its loss in local-projector overlap (Proposition 4); Theorem 5 bounds subspace change when source coupling is small relative to the block gap. These results characterize geometry, not prediction accuracy. Empirically, Fusion exceeds balanced covariance PCA with local residuals (48/16) by +1.29/+1.44 pp in AnyGraph/GraphFM. Split minus that capacity-matched covariance control is −1.80/+0.28 pp, so projector aggregation alone does not explain the gains. The earlier three-pipeline consensus-only variant loses 15.13 pp relative to Fusion, supporting the importance of retaining local information (Section 2; Appendices D.4 and F.2).

Q1 / Q3(a). Why orient the basis, and does orientation improve prediction?

Orientation fixes the coordinates supplied to the GFM after subspace selection. Procrustes maps the selected basis toward a common source reference while preserving its projector and within-graph Gram matrix (Proposition 6). Different graph-specific rotations can nevertheless produce different outputs from a shared encoder. GraphFM's graph-specific input MLP can absorb rotations at the function-class level, although training can differ. Orientation therefore supplies a coordinate convention, whose predictive value is an empirical question.

Our new 2×2 study crosses λ=0/1 with orientation off/on in AnyGraph and GraphFM: five targets, three seeds and five shot budgets. At λ=0, we retain Fusion's local PCA frame directly; orientation-on still uses the source reference. The four arms use the same interface conventions.

Mean macro-F1 change (pp)AnyGraphGraphFM
Add source-aware selection, orientation off+4.90+3.00
Add orientation, local selection (λ=0)+4.24+2.33
Add source-aware selection, orientation on+0.58−0.04
Add orientation, fused selection (λ=1)−0.09−0.71
Selection × orientation interaction−4.33−3.04

Selection and orientation each raise the mean when added to the local, unoriented interface. Their conditional gains shrink when the other component is present, producing the negative estimated interactions above. The two predictive effects therefore cannot be added independently. At λ=1, adding orientation also changes means by −0.39/−0.38/−1.52/−0.49 pp in MF-GIA/MDGFM/GraphGlue/BRIDGE; the earlier GraphLoRA study instead finds +1.16 pp. The six nonpositive mean effects have target-bootstrap intervals crossing zero. We therefore treat orientation as a common coordinate convention with architecture-dependent predictive effects, rather than a necessary source of Fusion's gains. The main configuration remains λ=1/on; these comparisons include source retraining (Appendix C.2, Tables 10 and 13, Appendix D.4.1).

Q2 / Q3(b). Equal weights and the possibility of negative transfer

We agree that equal numerical range does not imply equal predictive reliability. Both directional-support terms lie in [0,1], making λ=1 a fixed reference at a comparable numerical scale. Source informativeness can still differ across targets.

We varied λ over {0, 0.25, 1, 4} with orientation fixed. Relative to λ=1, changes at λ=0/0.25/4 are −0.58/−0.23/+0.59 pp in AnyGraph and +0.04/+0.05/−2.89 pp in GraphFM. AnyGraph shows mixed target effects; GraphFM declines on all five targets at λ=4. Stronger emphasis on source evidence can therefore hurt transfer, and its effect depends on the pipeline and target. We retain λ=1 and use the sweep to characterize this tradeoff. Appendix D.4.2 gives results at each weight for every target.

W2. Feature correspondence and independent column permutations

We have added theory distinguishing coordinate changes from correspondence changes. SoLIFT uses positional correspondence, and Procrustes does not infer feature semantics. Under exact spectral operators and strict selection gaps, a common permutation of the embedded feature coordinates preserves Fusion's node Gram matrix. For independent permutations, the bound relates subspace and Gram changes to consensus mismatch ε and fused eigengap δ: when ε<δ/2, the projector change is bounded by 2ε/δ. This is a conditional geometric guarantee; arbitrary independent permutations can alter correspondence and do not guarantee predictor invariance (Appendix H).

We tested source-only, target-only and combined permutations on five targets at 5-shot, using three seeds, three draws and paired episodes. Target-only tests reused the source checkpoints; source-only and combined tests refitted the interface and retrained the backbone. Fusion's signed mean changes are −0.12/+0.39/−0.33 pp in AnyGraph and −0.28/−0.59/+0.21 pp in GraphFM. All are within 0.6 pp, while its lead over matched SVD remains +5.45 to +6.38 pp and +2.00 to +2.80 pp, respectively. This supports average predictive robustness, alongside target variation: target-only permutation changes AnyGraph by +4.26 pp on Computers and −3.11 pp on Physics, and GraphFM by −3.32 pp on Computers. The experiment complements the conditional theory without establishing its small-mismatch premise for every graph (Appendix F.3).

W3. Within-pipeline comparisons, training policies and SVD correction

The primary evidence is the paired interface comparison within each pipeline. Each pair shares source graphs, training budget, checkpoint-selection policy, target adaptation and support/query episodes. GraphGlue's five-epoch schedule applies to both interfaces. Table 3 summarizes budgets, model selection and adaptation. Appendix B.4 describes implementation choices, and Appendix G gives computational resources. Cross-pipeline results assess whether the interface effect recurs across architectures, rather than rank the published systems.

With the corrected SVD implementation and matched re-evaluations, the mean Fusion–SVD gap is +2.64 pp, compared with +1.97 pp in the original submission. GraphGlue accounts for most of this change: its SVD mean changes from 37.04% to 32.52%, while Fusion remains at 37.47%. Fusion improves mean macro-F1 in all seven pipelines and, after averaging pipelines, at every shot budget. Accuracy, the original primary metric, also improves by +3.03 pp. Appendices C.1 and E provide the paired results and numerical implementation details.

W4. Quantitative neighborhood agreement and the role of t-SNE

We agree that independent t-SNE panels cannot quantify global geometry. We instead emphasize neighborhood agreement in the normalized 256-dimensional classifier-input space, before t-SNE: the same-class fraction among ten nearest query neighbors, averaged within class and then equally across classes. Fusion exceeds SVD in all three BRIDGE–Cora seeds, with mean agreement of 65.9% versus 62.1%.

This diagnostic uses the source-selected checkpoints from the original submission, including its original SVD implementation. Figure 3 now identifies the retrospectively selected 5-shot, seed-0 case and its 681 shared queries. The panels illustrate local organization; 10-NN quantifies consistency in the original representation space. This selected case concerns the complete trained model, rather than an isolated component or global geometry (Appendix D.6).

Reviewer_wV6EOpenReview ↗

Response

I thank the authors for their detailed and thoughtful rebuttal, which effectively addresses most of my concerns and provides valuable clarifications. I will maintain my score.

We appreciate your careful consideration of our rebuttal and your acknowledgment that it addresses most of your concerns. In light of the additional evidence and clarifications, we would be grateful if you could consider whether the revised manuscript merits a more positive assessment and a higher score.

We would also appreciate knowing which remaining concern is most important to your assessment, so that we can provide a focused clarification.

Reviewer_eqyeOpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a new approach to make node features compatible (e.g., in terms of dimensionality and meaning) when a single model must operate across different graphs. The problem is that standard dimensionality reduction methods do not account for directions that are not dominant within an individual dataset but recur across many source graphs. The authors propose SoLIFT, which first constructs a consensus subspace from the source graphs by averaging the projectors associated with their principal directions. One variant, called Split, reserves some coordinates for the shared directions and uses the remaining ones for local residual PCA. The other variant, Fusion, jointly selects all directions, balancing local information with agreement across the sources; it then aligns the basis to a common reference using a Procrustes transformation. SoLIFT is evaluated by integrating it into seven Graph Foundation Model pipelines and comparing it with PCA and several SVD variants. Fusion appears to be the most convincing variant.

Critical strengths

The method is well explained, motivated, and evaluated. The issue it addresses, ie, that the feature interface can lose information even before the main model comes into play, is highly relevant to GFMs. The comparison across seven pipelines makes the results less dependent on a single architecture.

Critical weaknesses
  • Probably the weakest point is that features are compared based on their column positions, even though their alignment is not guaranteed. It would be useful to develop an alignment based on feature semantics, rather than relying solely on their positions.

  • The experiments do not make it clear to me how reliable the method is when the source and target are different.

Questions for authors

It would be interesting to quantify the sensitivity of Split and Fusion to independent permutations of the features in the source graphs and the target graph.

1 comment
AuthorsOpenReview ↗

Thanks for your comment!

We appreciate your careful review and your recognition that the feature interface can lose information before the GFM encoder is applied. Your questions about feature correspondence and transfer across graphs motivated the additional theoretical analysis, permutation tests and stricter transfer evaluation below.

W1 / Q1. Feature semantics and column permutations

We have added a theoretical analysis that distinguishes a common change of coordinates from independent feature reorderings (Appendix H). With exact spectral operators and strict spectral gaps, applying the same permutation to every graph in the common embedded space preserves Fusion's node Gram matrix. For independent permutations, the subspace and Gram-matrix stability bounds apply when the induced source-consensus mismatch is smaller than half the fused eigengap. These results describe geometric invariance or conditional stability; they do not imply unchanged predictions under arbitrary independent permutations.

We tested the independent case on five targets at 5-shot, using three model seeds, three permutation draws and paired support/query episodes. Target-only permutations keep the source checkpoints fixed; source-only and combined permutations refit the interface and retrain the backbone. Mean macro-F1 changes from unpermuted features are shown in the order source only / target only / both:

MethodAnyGraph (pp)GraphFM (pp)
Fusion−0.12 / +0.39 / −0.33−0.28 / −0.59 / +0.21
Split−0.83 / +0.38 / +2.26−0.15 / +0.76 / +1.24

Fusion's signed mean change stays within 0.6 pp, and its average advantage over matched SVD remains positive in every condition: +5.45 to +6.38 pp in AnyGraph and +2.00 to +2.80 pp in GraphFM. Individual targets remain more sensitive: target-only permutation changes AnyGraph Fusion by +4.26 pp on Computers and −3.11 pp on Physics, and GraphFM Fusion by −3.32 pp on Computers. Source-only permutation also reduces AnyGraph Split on Cora by 7.08 pp. The study supports average predictive robustness for Fusion alongside target-dependent sensitivity (Appendix F.3).

We agree that semantic alignment is a feasible extension when feature names or descriptions are available. Such information could establish cross-graph correspondence before SoLIFT's source–local compression. The current positional convention supplies a numerical correspondence, and Procrustes aligns coordinate frames after subspace selection; neither infers feature semantics. We plan to investigate this semantic alignment stage in future work.

W2. Transfer to unseen graphs

We directly tested transfer without citation/coauthorship source graphs through a strict bibliographic-family holdout in AnyGraph. Citeseer, PubMed, arXiv and Coauthor-CS were excluded from source-interface fitting, backbone pretraining and source-validation checkpoint selection. Both methods were freshly trained on the same five remaining sources. Under this shared holdout setting, Fusion exceeds SVD by +7.86 pp on Cora, +11.37 pp on Coauthor-Physics and +4.39 pp on DBLP. The three-target mean is +7.88 pp after averaging three seeds and five shot budgets. This provides direct evidence that the complete Fusion interface can retain gains without sources from the held-out family. The exploratory study covers one predefined family in AnyGraph (Appendix I).

Under the original nine-source setting, the matched comparison now includes DBLP, Minesweeper and Questions, extending AnyGraph and GraphFM to eight targets. Fusion gains +3.45 pp on average in AnyGraph (4/8 targets) and +1.43 pp in GraphFM (6/8). DBLP improves by +10.02 and +2.27 pp, respectively. Across the additional three targets, the means are −0.89 and +0.66 pp, with declines on Questions in both pipelines. The broader collection therefore retains positive overall gains while revealing target-specific transfer limitations (Appendix F.2).

To assess architectural coverage, we also compared Fusion with SVD across all seven pipelines on the original five unseen targets. Fusion improves each pipeline's mean macro-F1, averaging +2.64 pp across pipelines, and the aggregate gain is positive at every shot budget. Accuracy, the original primary metric, improves by +3.03 pp on the same runs. These comparisons use graph-held-out, transductive evaluation: target preprocessing can use the unlabeled target graph, while query labels are used only for evaluation (Appendix C.1).

Comments on the paper

6 comments
AuthorsOpenReview ↗

Thanks for your comment!

We appreciate your careful assessment of the method and your concrete suggestions for testing feature correspondence, transfer beyond related graphs, and the source-consensus mechanism. We have added the requested controls and a strict family holdout, and clarified how the common/local allocation was selected.

W1 / Q1. Independent feature-column permutations

We added a theoretical analysis of coordinate changes and the requested independent-permutation experiments. The analysis distinguishes a common relabeling of coordinates from a change in correspondence between graphs. With exact spectral operators and spectral gaps, applying the same permutation to every graph in the embedded space preserves Fusion's node Gram matrix. Under independent permutations, Appendix H bounds the change in the selected subspace and node Gram matrix when the induced consensus mismatch is below half the fused eigengap. This identifies a condition for geometric stability; it does not assume that arbitrary independent permutations preserve cross-graph correspondence.

We then tested source-only, target-only and combined permutations on five targets at 5-shot, using three model seeds, three permutation draws and paired support/query episodes. Each graph's native feature columns were independently permuted. Target-only tests kept the source checkpoint fixed; source-only and combined tests refitted the interface and retrained the backbone. The table reports signed mean macro-F1 changes from native features, in that order.

MethodAnyGraph (pp)GraphFM (pp)
Fusion−0.12 / +0.39 / −0.33−0.28 / −0.59 / +0.21
Split−0.83 / +0.38 / +2.26−0.15 / +0.76 / +1.24

Fusion's mean changes stay within 0.6 pp, and its average advantage over matched SVD persists in every condition: +5.45 to +6.38 pp in AnyGraph and +2.00 to +2.80 pp in GraphFM. Thus the tested changes in feature ordering do not remove its average benefit. Individual targets are more sensitive: target-only permutation changes AnyGraph Fusion by +4.26 pp on Computers and −3.11 pp on Physics, and GraphFM Fusion by −3.32 pp on Computers. Source-only permutation lowers AnyGraph Split on Cora by 7.08 pp (Appendix F.3).

The experiments measure predictive sensitivity of the implemented pipelines, while the theory characterizes geometry under exact spectral operators. SoLIFT still uses positional correspondence; projector aggregation removes arbitrary basis choices within a PCA subspace but does not infer feature semantics. Feature names or descriptions could establish correspondence before source–local compression, which we regard as a promising extension.

W2–W3 / Q2. Transfer beyond related source graphs and broader target coverage

In AnyGraph, Fusion retains a positive advantage when the entire citation/coauthorship family is excluded from source training. Following your suggestion, we predefined this bibliographic family and removed Citeseer, PubMed, arXiv and Coauthor-CS from interface fitting, backbone pretraining and source-validation checkpoint selection. Both Fusion and SVD were freshly trained on WikiCS, Amazon-ratings, Amazon-photo, Actor and Chameleon. Family membership is defined by graph relations, so WikiCS remains a web graph.

We evaluated Cora, Coauthor-Physics and DBLP with three seeds and five support budgets (1/3/5/10/20 labels per class), pairing support/query episodes across methods:

Held-out targetFusion minus SVD, macro-F1 (pp)
Cora+7.86
Coauthor-Physics+11.37
DBLP+4.39
Equal-target mean+7.88

All three targets improve. This directly supports transfer of the complete Fusion interface without source graphs from the held-out family. It is an exploratory study of one predefined family in AnyGraph; extending this test to further families and architectures would establish its breadth more fully (Appendix I).

We also expanded the matched AnyGraph/GraphFM benchmark from five to eight targets by adding DBLP, Minesweeper and Questions. Fusion's average gain remains positive in both pipelines:

PipelineEight-target gain (pp)Targets improvedExploratory 95% target-bootstrap interval
AnyGraph+3.454/8[−3.06, +10.74]
GraphFM+1.436/8[+0.28, +2.62]

DBLP improves by +10.02/+2.27 pp in AnyGraph/GraphFM. Across the three additional targets, the means are −0.89/+0.66 pp, with declines on Questions in both pipelines. The expanded benchmark therefore supports an overall benefit while identifying domains where transfer is less effective (Appendix F.2).

The architectural evidence is complementary: on the original five targets, Fusion improves mean macro-F1 in all seven pipelines and at every shot budget after averaging pipelines. The overall gain is +2.64 pp; removing any one target leaves a positive mean of +1.16 to +3.55 pp. Accuracy, the original primary metric, improves by +3.03 pp on the same runs, with positive aggregate effects on all five targets (Appendix C.1).

The five-target macro-F1 interval remains [−0.06, +5.94] pp, so the small domain sample still limits broad generalization claims. The expanded target collection and family holdout add complementary evidence: gains recur across architectures and persist after excluding related sources. We therefore interpret the results as average improvements in these tested settings, with target-specific variation. Evaluation remains transductive, using full unlabeled target graphs while reserving query labels for evaluation.

Q3. Simpler source-informed PCA controls

We added the suggested pooled and graph-balanced source PCA bases, together with a capacity-matched common/local control. All source covariances use the same graphwise centering. Pooled PCA weights graphs by node count; graph-balanced PCA weights graphs equally. Each produces a fixed rank-64 source basis. The third control uses 48 balanced-PCA directions plus 16 residual local-PCA directions, matching Split's allocation and residual construction.

Fusion minus control, eight-target mean (pp)AnyGraphGraphFM
Fixed pooled PCA+11.11+16.52
Fixed balanced PCA+10.78+17.36
Balanced PCA with residuals (48/16)+1.29+1.44

The fixed bases keep source-fitted directions and use the source-supported feature width. Local-residual methods adapt to each graph and can retain wider target columns, so this comparison includes both local adaptation and column coverage. The 48/16 control addresses a more specific question: whether the common-space estimator improves performance when local capacity is held fixed.

Replacing balanced PCA with projector aggregation in Split changes the eight-target mean by −1.80 pp in AnyGraph and +0.28 pp in GraphFM. Projector aggregation alone therefore has mixed effects. Fusion improves over this capacity-matched control by +1.29/+1.44 pp; in GraphFM, it improves on six of eight targets. This supports the full source–local interface, which jointly selects directions and orients their coordinates. The baselines help separate this benefit from a claim that the projector estimator itself must outperform covariance PCA (Appendix F.2).

Q4. How 48/16 was selected and how to choose the allocation from sources

The original 48/16 allocation was chosen after favorable early target results and then fixed across all seven pipelines. The retrospective capacity analysis did not constitute source-only selection. We have made this history explicit and added a source-data procedure for selecting the allocation.

In AnyGraph, we reserved Citeseer, Amazon-photo and Actor as source-development graphs, excluding them from interface fitting, backbone training and checkpoint selection. On the other six sources, we trained five allocations with common dimensions 0/16/32/48/64 at total width 64. Selection used mean validation macro-F1 on the three development graphs, with three seeds and matched 5-shot episodes.

The development score selected 32/32 at 42.44% macro-F1, compared with 40.34% for 48/16. We then fixed 32/32 and refitted on all nine sources with three seeds. Its descriptive test-set mean at 5-shot across the five original targets was 47.06%. This final number describes the selected model; the allocation decision used the source-development scores. The experiment provides a practical way to choose common/local capacity from source data (Appendix D.3.1, Tables 11–12).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.