Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment

Bokui Wang; Boyan Deng; Shuo Wang; Zhao Kang; Zhixiang Shen

REVIEW 2 major objections 5 minor 2 cited by

Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment

T0 review · 2 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that aligning graph topologies across source domains lets a single pre-trained model transfer to unseen citation, web, and social graphs, beating multi-domain baselines by up to 18.04% in one-shot node classification.

desk verdict Empirically solid few-shot transfer paper; the theory section is borrowed and the key invariance claim is asserted rather than shown. read the letter →

arxiv 2502.02017 v2 pith:I42NUUGW submitted 2025-02-04 cs.SI cs.AIcs.LG

Shuo Wang , Bokui Wang , Zhixiang Shen , Boyan Deng , Zhao Kang This is my paper · ORCID

classification cs.SIcs.AIcs.LG

keywords graphfoundationmodelsmulti-domainpre-trainingtopologyalignmentstructurelearningdomaingeneralizationfew-shotnodeclassificationheterophilyprompttuning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Multi-domain graph pre-training faces a basic obstacle: citation networks, web graphs, and social networks organize their edges differently, so a model that memorizes one topology transfers poorly. This paper claims that explicitly aligning topologies across domains fixes that. The proposed MDGFM framework projects all graphs into a shared semantic space, then refines each adjacency matrix with a graph-structure-learning step that is meant to strip away noise and domain-specific wiring while keeping task-relevant structure. A contrastive objective ties the refined graph to the original, and a dual-prompt mechanism adapts the frozen encoder to an unseen target with very few labels. On one-shot and few-shot node classification over homophilic and heterophilic benchmarks, the paper reports consistent gains over the strongest multi-domain baselines, up to 18.04% in the one-shot setting, plus resilience to edge attacks.

What carries the argument

The load-bearing object is the graph-structure-learning (GSL) refinement module, which rebuilds each adjacency matrix from a balance-token fusion of raw features and one-hop aggregated features, then applies kNN sparsification, symmetrization, ReLU activation, and normalization. The paper identifies this module with the 'optimal invariant graph learner' $\Phi^*(G)$ of Assumption 5.2: it is supposed to remove domain-specific variation while preserving sufficient task-relevant information. The other machinery is the two-view contrastive loss $\mathcal{L} = -I(G_{i1}; G_{i2} \dagger I_e) - I(G_{i1}; G_{i2} \dagger A'_i)$, where the refined adjacency supplies additional positive pairs, and the dual-prompt adapter that keeps the encoder frozen.

What would settle it

Measure the cross-domain gap in edge patterns (e.g., homophily ratio or spectral profile) between refined adjacency matrices; if refinement does not bring two source domains' topologies closer than the originals were, the claimed alignment mechanism is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that topological information, not just node features, is the transferable currency across graph domains, and that a learned graph refinement can serve as the 'optimal invariant graph learner' that the theory requires. Features from different domains are first projected to a common dimension and modulated by domain tokens and a shared token, producing a unified semantic space. A balance token then fuses each node's features with its aggregated neighborhood, and a kNN-based post-processing step yields a sparsified, symmetrized, normalized refined adjacency $A'_i$. Pre-training maximizes mutual information between the original view $(A_i, X'_i)$ and the refined view $(A'_i, X'_i)$, which the paper argues preserves task-relevant information while making the representation domain-invariant. At test time, a meta prompt and a specific prompt adapt the frozen encoder, and the final representation is $Z = GE(A'_T, \beta p_m(X_T)+(1-\beta)p_s \odot X_T; \theta_{pre})$. The reported result is that this pipeline outperforms multi-domain baselines on one-shot and few-shot node classification on both homophilic and heterophilic unseen graphs and stays stable under topology attacks.

Load-bearing premise

The method's guarantees depend on the assumption that its graph-refinement step really removes domain-specific variation while keeping the information needed for classification, and this invariance is asserted rather than proven or directly measured.

Editorial extensions

If this is right

If MDGFM is right, a single pre-trained graph encoder can serve citation, web, and social-network domains, with downstream adaptation limited to a prompt and a few label samples.
Topology refinement substitutes for task-specific architecture choices: the same frozen encoder handles both homophilic and heterophilic graphs because the refined adjacency encodes the relevant wiring.
Because the refined graph is used as an extra contrastive view, noisy or adversarial edges are down-weighted during pre-training, which is the stated source of the model's robustness under edge deletions, additions, and meta-attacks.
The prompt coefficients and the specific prompt are the only parameters trained downstream, so adaptation cost stays low even when the target domain is large-scale, as shown on the 41,554-node Penn94 graph.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the paper's own guarantee rests on identifying GSL with the invariant learner, but it never directly measures invariance of the refined graphs; a cross-domain comparison of refined adjacency matrices (e.g., homophily ratio or spectral distance before vs. after refinement) would test this premise directly.
Editorial inference: because kNN sparsification is the core of the refiner, the reported low sensitivity to $k$ may not survive on graphs with very different degree distributions; an adaptive per-domain $k$ rule is a natural extension.
Editorial inference: the method's success on heterophilic graphs suggests that topology alignment can serve as a replacement for heterophily-specific architectures; one could test this by applying MDGFM to a heterophily benchmark suite with varied homophily ratios.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the empirical contribution is real and worth engaging; the theory section is borrowed and the load-bearing claim that the GSL module implements an invariant graph learner is asserted, not shown.

What is new is the specific assembly: adaptive balance tokens that mix features and aggregated topology, a kNN-based graph structure learning step applied to every domain, and a dual prompt (meta plus specific) for transfer. That combination is not in GCOPE or MDGPT, and the evaluation is more thorough than typical for this subfield: seven datasets including large-scale Penn94, homophilic and heterophilic targets, 1-shot and K-shot settings, ablations of each component, a source-domain sensitivity analysis, and robustness to random edge edits and meta-attack. The gains relative to the second-best method are consistent and reach 18.04% in one-shot. The main tables report mean and std over five runs, which is more than many graph learning papers do.

The soft spot is Section 5. Theorem 5.1 is from Albuquerque et al. (2019). Assumption 5.2 postulates an optimal invariant learner Φ*, and the paper then states that Φ* 'is implemented as GSL procedure on all domains.' Nothing in the loss or the structure-learning procedure enforces the invariance property—the contrastive objective maximizes agreement between original and refined views, which can preserve domain-specific structure—and no experiment checks whether the refined graphs are actually invariant across source domains. The sufficiency claim is also just asserted. The independence and product-form conditions of Theorem 5.3 are never argued for this pipeline. So the theoretical guarantee is decorative. That matters because the paper's framing attributes the gains to topology alignment as invariant learning; they could instead come from the extra tokens or from kNN denoising. I would not call this fatal—the experiments stand on their own—but it does mean the theory section should either be substantiated or reframed as motivation.

Minor: the ablation and robustness figures lack error bars, and the code link is given but not verifiable from the text.

Who should read this: researchers working on multi-domain graph pretraining and few-shot node classification. The prompt design and the stability analysis under source-domain removal are genuinely useful. I would send it to review; the empirical contribution deserves referee time, and the theory can be revised or cut down to an honest discussion. My main request as a referee would be to either prove or drop the invariant-learner identification and instead measure domain discrepancy before and after GSL.

Referee Report

2 major / 5 minor

Summary. The paper proposes MDGFM, a multi-domain graph foundation model that unifies source graphs in a projected feature space, refines each graph with a graph structure learning (GSL) module, and transfers to unseen target domains via a dual-prompt strategy. The empirical section reports one-shot and few-shot node classification results on six benchmark graphs plus Penn94, with ablations, domain-sensitivity analysis, and robustness tests under edge perturbations and Metattack. The theoretical section claims domain generalization guarantees by identifying the GSL module with an optimal invariant graph learner from prior work.

Significance. If the empirical results hold, the paper makes a useful contribution to multi-domain graph pre-training: it is, to my knowledge, the first to explicitly combine topology refinement with prompt-based transfer across homophilic and heterophilic source domains, and the gains over GCOPE and MDGPT on Cornell, Squirrel, and Chameleon are non-trivial. The authors provide a public code repository, report seven datasets, include component ablations, and examine robustness under two attack families. The theoretical section, however, does not derive a guarantee for the proposed pipeline; it imports bounds from Albuquerque et al. (2019) and Li et al. (2022) and then asserts, rather than proves, that the GSL module is the invariant learner required by Assumption 5.2. This gap is load-bearing because the paper's central narrative is that topology alignment implements invariant graph learning, and it is the main reason I cannot recommend acceptance without revision.

major comments (2)

[Section 5, Assumption 5.2 and the paragraph following it] The theoretical guarantee rests on identifying the GSL module with the optimal invariant graph learner Φ*: the text states that Φ* is 'implemented as GSL procedure on all domains in our work' and that maximizing mutual information 'retains sufficient task-relevant information.' Neither claim is established. The GSL module in Section 4.2 (Eq. 3, Appendix F) computes a kNN graph from Hi = tBi ⊙ [X'_i, A_i^r X'_i] followed by symmetrization and normalization; it contains no term that penalizes differences in P(Y | Φ*(G)) across domains, and no experiment or quantitative check shows that refined adjacency matrices A'_i are invariant across source domains. Similarly, the contrastive loss in Eqs. (13)-(15) is an InfoNCE-style lower bound on mutual information between node representations from two views, which does not by itself imply that the refined graph preserves label-relevant information. As a result, Theorem 5.3's conclusion—that the proposed predictor attains the minimal error bound—is asserted rather than derived. I ask the authors to either prove that their objective enforces Assumption 5.2, add a direct invariance metric on the refined graphs, or substantially weaken the theoretical claims to match what is actually shown.
[Section 5, Theorem 5.3 and conditions (1)-(2)] Theorem 5.3 requires conditions that are not checked or argued for the proposed pipeline: (1) Φ*(G) is independent of the variant subgraph G \ Φ*(G), and (2) there exists an environment e' under which the joint distribution factorizes as P^{e'}(G, Y) = P^{e'}(Φ(G), Y) P^{e'}(G \ Φ(G)) and the marginal P^{e'}(Φ(G)) equals the source marginal P^e(Φ(G)). The paper gives no reason to believe that the kNN-refined graph produced by Eq. (3) and Appendix F satisfies independence or this product-form condition, particularly for heterophilic graphs where the refined structure depends on both features and the original adjacency. Because Theorem 5.3 is the only result that connects the method to the optimal invariant predictor, the absence of these checks is a load-bearing gap. The authors should either verify these conditions on the actual model or recast Section 5 as a motivation for the design rather than a formal guarantee.

minor comments (5)

[Section 4.2, Eq. (4)] The notation '†' in Eq. (4) is undefined in the main text. Appendix G later clarifies that the two terms use I_e and A'_i to define positive samples, but the main-text formula is confusing and should be rewritten with explicit definitions.
[Section 6.3, Figure 2] The ablation bar chart reports only point estimates without error bars, even though Tables 1 and 2 show that the one-shot and few-shot results carry standard deviations of 4-13 points. Without error bars or a significance test, claims like 'w/o-sumtoken exhibits the weakest performance across most datasets' are difficult to assess.
[Section 6.5, Figure 3 and Appendix D.1, Figure 4] The robustness curves show no error bars or confidence intervals. Given the high variance in the one-shot results reported in Table 1, the claim that MDGFM is 'quite stable' under attacks would be strengthened by reporting variance across the five runs.
[Table 4] The baseline column is labeled 'SAMGPT' in Table 4 but 'MDGPT' in the main text and in the methods description. Please use one consistent name for the method (the reference list cites it as 'SamGPT').
[Appendix B] The last sentence of Appendix B says 'We will include this complexity analysis in the final version for completeness,' which is a remnant of the submission process. Since the appendix is already present in this version, the sentence should be removed or rephrased.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: empirical results are benchmarked externally and the theoretical claims import external theorems; the GSL-as-Phi* identification is an unproven assumption, not a by-construction reduction.

full rationale

The paper's central empirical claims are tested on held-out target domains against external baselines (GCN, GAT, DGI, GraphCL, GPPT, GPF, GCOPE, MDGPT) using a common backbone and fixed few-shot protocol. No reported metric is a fitted value or a relabeled input, so the 'fitted input called prediction' pattern does not apply. The theory is not self-referential in a load-bearing way: Theorem 5.1 is taken from Albuquerque et al. (2019) and Theorem 5.3 from Li et al. (2022), neither of which is authored by the present group. Self-citations to the authors' earlier GSL/heterophily work appear only as building blocks (e.g., 'Following the approach proposed in previous works (Liu et al., 2022; Shen et al., 2024)' for kNN sparsification and post-processing), and the central transfer result does not reduce to those citations. The genuine weakness is Section 5's assertion that Phi*(G) 'is implemented as GSL procedure on all domains in our work' without proving that the kNN-based refinement of Eq. (3)/Appendix F satisfies the invariance and sufficiency properties of Assumption 5.2, or that Theorem 5.3's product-form conditions hold. This is an unsupported identification and a rigor gap, not a circular reduction: the conclusions are not equivalent to the inputs by construction, and no parameter is fitted to the target and then renamed a prediction. The evaluation is self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

The method uses many learned parameters and per-dataset hyperparameters, which are the price of the approach. The key non-standard axiom is the assumed existence and realizability of the invariant graph learner Phi* by the GSL procedure; this is the main unproven input. Standard math (MI lower bound) and PCA assumption round out the ledger.

free parameters (9)

Domain tokens t_Di per source domain = learned during pre-training
Element-wise filters on features to encode domain semantics (Eq. 2).
Shared token t_S = learned
Common semantic anchor applied across domains (Eq. 2).
Balance token t_Bi per domain = learned
Weights concatenation of features and aggregated topology in Eq. (3).
Meta-prompt coefficients alpha_1..alpha_N = learned
Mixing weights for source domain tokens in meta prompt Eq. (5).
Specific prompt p_s = learned on target
Task-specific prompt for downstream domain in Eq. (6).
Blending coefficient beta = learned
Balances meta prompt and specific prompt in Eq. (6).
kNN neighborhood size k = 30 homophilic, 15 heterophilic
Set per dataset type for graph structure learning; sensitivity shown in Appendix D.2.
Unified feature dimension d = 50
PCA output dimension fixed across domains.
Contrastive temperature tau_c = 0.2
Fixed in Appendix G.

assumptions (4)

domain assumption Covariate shift: target feature distribution is approximated by a convex combination of source feature distributions.
Invoked in Section 5 before Theorem 5.1 to frame the generalization bound.
ad hoc to paper There exists an optimal invariant graph learner Phi* satisfying the invariance and sufficiency properties of Assumption 5.2.
Assumed in Section 5; the paper asserts its GSL module implements Phi* without proof.
standard math The sample-based mutual information lower bound (Liang et al. 2023) is valid for the contrastive objective.
Used in Appendix G to justify the loss computations.
domain assumption PCA projection to 50 dimensions retains sufficient cross-domain semantic information.
The projection in Eq. (1) assumes principal components are the right shared basis; no analysis of information loss is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment." pith.science (2026). https://pith.science/paper/I42NUUGW

@misc{pith2026250202017,
  author       = {Pith},
  title        = {Pith review of: Multi-Domain Graph Foundation Models: Robust Knowledge Transfer via Topology Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I42NUUGW}},
  note         = {Machine review of arXiv:2502.02017}
}

read the original abstract

Recent advances in CV and NLP have inspired researchers to develop general-purpose graph foundation models through pre-training across diverse domains. However, a fundamental challenge arises from the substantial differences in graph topologies across domains. Additionally, real-world graphs are often sparse and prone to noisy connections and adversarial attacks. To address these issues, we propose the Multi-Domain Graph Foundation Model (MDGFM), a unified framework that aligns and leverages cross-domain topological information to facilitate robust knowledge transfer. MDGFM bridges different domains by adaptively balancing features and topology while refining original graphs to eliminate noise and align topological structures. To further enhance knowledge transfer, we introduce an efficient prompt-tuning approach. By aligning topologies, MDGFM not only improves multi-domain pre-training but also enables robust knowledge transfer to unseen domains. Theoretical analyses provide guarantees of MDGFM's effectiveness and domain generalization capabilities. Extensive experiments on both homophilic and heterophilic graph datasets validate the robustness and efficacy of our method.

Figures

Figures reproduced from arXiv: 2502.02017 by the authors.

**Figure 1.** The overall framework of the proposed MDGFM. metrics like homophily and heterophily edge ratios (Zhu et al., 2020). These structural characteristics underscore the importance of aligning graph structure patterns to enable meaningful knowledge transfer (Sun et al., 2022b; Zheng et al., 2022). Rather than solely focusing on direct structural unification, our approach seeks to synchronize more reliable topology inform… view at source ↗

**Figure 2.** Ablation studies on key components. in the one-shot scenario. Notably, on heterophilic graphs, certain methods exhibit a decline in performance compared to the one-shot case. This degradation is primarily due to the introduction of noise from the few-shot samples. In contrast, our model remains robust and does not experience performance deterioration as the number of training samples increases. Quantitatively, acros… view at source ↗

**Figure 3.** Performance of robustness analysis, where the dataset in parentheses represents the target domain [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Performance of robustness analysis, where Cora is the target domain. (a) Homophilic graphs (b) Heterophillic graphs [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

**Figure 5.** Figure 5: Sensitivity study of k. homophilic or heterophilic. When the target domain is homophilic graphs, larger values of k (such as 20 or 30) yield the best results. In contrast, for heterophilic graphs, relatively smaller values of k can produce satisfactory outcomes. E. Exp…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Beyond Feature and Structure Alignment: Learning Transferable Propagation Knowledge for Graph Foundation Models
cs.LG 2026-07 conditional novelty 6.0 of 10

ProGFM transfers graph knowledge across domains by learning a prototype bank of per-edge, per-dimension propagation strengths and using them to modulate message passing on unseen graphs.
Unified Multi-Domain Graph Pre-training for Homogeneous and Heterogeneous Graphs via Domain-Specific Expert Encoding
cs.LG 2026-02 conditional novelty 5.0 of 10

GPH^2 pre-trains one expert per graph on edge-dropped or meta-path views and fuses frozen experts with class-wise attention, outperforming type-specific graph pre-training baselines.

Reference graph

Works this paper leans on

14 extracted references · 7 canonical work pages · cited by 2 Pith papers

[5]

Towards unsupervised deep graph structure learning

Liu, Y ., Zheng, Y ., Zhang, D., Chen, H., Peng, H., and Pan, S. Towards unsupervised deep graph structure learning. In Proceedings of the ACM Web Conference 2022 , pp. 1392–1403,

2022
[8]

Samgpt: Text-free graph foundation model for multi-domain pre- training and cross-domain adaptation

Yu, X., Gong, Z., Zhou, C., Fang, Y ., and Zhang, H. Samgpt: Text-free graph foundation model for multi-domain pre- training and cross-domain adaptation. In THE WEB CON- FERENCE 2025,

work page 2025
[9]

Multi-domain knowledge graph collaborative pre-training and prompt tuning for diverse downstream tasks

Zhang, Y ., Hu, B., Chen, Z., Guo, L., Liu, Z., Zhang, Z., Liang, L., Chen, H., and Zhang, W. Multi-domain knowledge graph collaborative pre-training and prompt tuning for diverse downstream tasks. arXiv preprint arXiv:2405.13085, 2024b. Zhao, H., Chen, A., Sun, X., Cheng, H., and Li, J. All in one and one for all: A simple yet effective method towards cr...

arXiv
[11]

Deep graph structure learning for robust representa- tions: A survey

Zhu, Y ., Xu, W., Zhang, J., Liu, Q., Wu, S., and Wang, L. Deep graph structure learning for robust representa- tions: A survey. arXiv preprint arXiv:2103.03036, 14: 1–1,

arXiv
[13]

The hyperparameters corresponding to each dataset. Target domainPre-training learning rateDownstream learning rateEpoch Unified dimensionDropout Downstreamk Cora 0.0075 0.001 60 50 0.1 30 Citeseer 0.001 0.001 60 50 0.1 30 Pubmed 0.0001 0.0015 60 50 0.1 30 Squirrel 0.01 0.0003 100 50 0.1 15 Chameleon 0.02 0.01 100 50 0.1 15 Cornell 0.02 0.0003 100 50 0.1 1...

work page 2022
[14]

to enhance the model’s scalability. This approach avoids the resource-intensive computation and storage of explicit similarity matrices, reducing the complexity from O(|V |2) to O(|V |B), where |V | is the number of nodes and B is the batch size of the sparse kNN. Symmetrization and Activation. Since real-world connections are typically bidirectional, we ...

work page 2024
[2010]

H., and Mitliagkas, I

Albuquerque, I., Monteiro, J., Darvishi, M., Falk, T. H., and Mitliagkas, I. Generalizing to unseen domains via distribution matching. arXiv preprint arXiv:1911.00804,

arXiv 1911
[2012]

The nodes are annotated with the reported gender of each user, and the objective is to predict this gender

dataset is a large-scale social network derived from the Facebook 100 networks of university students from 2005, where each node corresponds to an individual student. The nodes are annotated with the reported gender of each user, and the objective is to predict this gender. The features associated with the nodes include major, secondary major/minor, dorm/...

work page 2005

Show all 14 references

[2019]

A survey on in- context learning

Dong, Q., Li, L., Dai, D., Zheng, C., Ma, J., Li, R., Xia, H., Xu, J., Wu, Z., Chang, B., et al. A survey on in- context learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1107–1128,

2024
[2020]

Self-supervised learning on graphs: Deep insights and new direction

Jin, W., Derr, T., Liu, H., Wang, Y ., Wang, S., Liu, Z., and Tang, J. Self-supervised learning on graphs: Deep insights and new direction. arXiv preprint arXiv:2006.10141, 2020a. Jin, W., Ma, Y ., Liu, X., Tang, X., Wang, S., and Tang, J. Graph structure learning for robust g...

2006 arXiv
[2021]

Making pre-trained lan- guage models better few-shot learners

Gao, T., Fisch, A., and Chen, D. Making pre-trained lan- guage models better few-shot learners. arXiv preprint arXiv:2012.15723,

2012 arXiv
[2022]

Inductive graph alignment prompt: Bridging the gap between graph pre- training and inductive fine-tuning from spectral perspec- tive

Yan, Y ., Zhang, P., Fang, Z., and Long, Q. Inductive graph alignment prompt: Bridging the gap between graph pre- training and inductive fine-tuning from spectral perspec- tive. In Proceedings of the ACM on Web Conference 2024, pp. 4328–4339,

2024
[2023]

Xia, J., Wu, L., Chen, J., Hu, B., and Li, S. Z. Simgrace: A simple framework for graph contrastive learning without data augmentation. In Proceedings of the ACM Web Conference 2022, pp. 1070–1079,

2022
[2024]

S., and Pan, S

Zheng, X., Wang, Y ., Liu, Y ., Li, M., Zhang, M., Jin, D., Yu, P. S., and Pan, S. Graph neural networks for graphs with heterophily: A survey. arXiv preprint arXiv:2202.07082,

arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [5]

Towards unsupervised deep graph structure learning

Liu, Y ., Zheng, Y ., Zhang, D., Chen, H., Peng, H., and Pan, S. Towards unsupervised deep graph structure learning. In Proceedings of the ACM Web Conference 2022 , pp. 1392–1403,

2022

[2] [8]

Samgpt: Text-free graph foundation model for multi-domain pre- training and cross-domain adaptation

Yu, X., Gong, Z., Zhou, C., Fang, Y ., and Zhang, H. Samgpt: Text-free graph foundation model for multi-domain pre- training and cross-domain adaptation. In THE WEB CON- FERENCE 2025,

work page 2025

[3] [9]

Multi-domain knowledge graph collaborative pre-training and prompt tuning for diverse downstream tasks

Zhang, Y ., Hu, B., Chen, Z., Guo, L., Liu, Z., Zhang, Z., Liang, L., Chen, H., and Zhang, W. Multi-domain knowledge graph collaborative pre-training and prompt tuning for diverse downstream tasks. arXiv preprint arXiv:2405.13085, 2024b. Zhao, H., Chen, A., Sun, X., Cheng, H., and Li, J. All in one and one for all: A simple yet effective method towards cr...

arXiv

[4] [11]

Deep graph structure learning for robust representa- tions: A survey

Zhu, Y ., Xu, W., Zhang, J., Liu, Q., Wu, S., and Wang, L. Deep graph structure learning for robust representa- tions: A survey. arXiv preprint arXiv:2103.03036, 14: 1–1,

arXiv

[5] [13]

The hyperparameters corresponding to each dataset. Target domainPre-training learning rateDownstream learning rateEpoch Unified dimensionDropout Downstreamk Cora 0.0075 0.001 60 50 0.1 30 Citeseer 0.001 0.001 60 50 0.1 30 Pubmed 0.0001 0.0015 60 50 0.1 30 Squirrel 0.01 0.0003 100 50 0.1 15 Chameleon 0.02 0.01 100 50 0.1 15 Cornell 0.02 0.0003 100 50 0.1 1...

work page 2022

[6] [14]

to enhance the model’s scalability. This approach avoids the resource-intensive computation and storage of explicit similarity matrices, reducing the complexity from O(|V |2) to O(|V |B), where |V | is the number of nodes and B is the batch size of the sparse kNN. Symmetrization and Activation. Since real-world connections are typically bidirectional, we ...

work page 2024

[7] [2010]

H., and Mitliagkas, I

Albuquerque, I., Monteiro, J., Darvishi, M., Falk, T. H., and Mitliagkas, I. Generalizing to unseen domains via distribution matching. arXiv preprint arXiv:1911.00804,

arXiv 1911

[8] [2012]

The nodes are annotated with the reported gender of each user, and the objective is to predict this gender

dataset is a large-scale social network derived from the Facebook 100 networks of university students from 2005, where each node corresponds to an individual student. The nodes are annotated with the reported gender of each user, and the objective is to predict this gender. The features associated with the nodes include major, secondary major/minor, dorm/...

work page 2005