Pith. sign in

REVIEW 5 major objections 6 minor 19 references

Continuous graph-text alignment beats discrete token compression in federated graph foundation models, with gains up to 14.37%.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 06:12 UTC pith:AP72YT53

load-bearing objection FedGALA is a plausible new combination, but its central claim—continuous alignment beats VQ—is never tested directly; the paper needs a VQ control and code before the 14.37% number can be believed. the 5 major comments →

arxiv 2601.21369 v2 pith:AP72YT53 submitted 2026-01-29 cs.LG

Rethinking Federated Graph Foundation Models: A Graph-Language Alignment-based Approach

Yinlin Zhu , Di Wu , Xianzhi Zhang , Yuming Ai , Xunkai Li , Miao Hu , Guocong Quan This is my paper
classification cs.LG
keywords federated graph foundation modelsgraph-language alignmentcontrastive learningvector quantizationprompt tuningtext-attributed graphsfederated learninggraph neural networks
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that federated graph foundation models lose important information when they compress graph semantics into discrete tokens via vector quantization, and that this loss is avoidable. It proposes FedGALA, which aligns graph and text encoders in a continuous embedding space using unsupervised contrastive learning during federated pre-training, then adapts frozen backbones with lightweight prompt tuning. The central claim is that continuous alignment preserves transferable structural knowledge better than quantization, leading to the best results across eight datasets and three tasks, gains up to 14.37% over baselines, faster convergence, and lower communication costs. A sympathetic reader would care because it offers a simpler, bandwidth-friendly alternative to quantization-based federated graph foundation models without sacrificing privacy.

Core claim

FedGALA's central discovery is that the discrete vector quantization used by prior federated graph foundation models causes irreversible representational loss, and that continuous graph-text contrastive alignment avoids it. By decoupling the graph encoder into a shared structural encoder—the only part transmitted and aggregated—and a local semantic encoder, and aligning their fused output with a frozen language model's text embeddings through a symmetric contrastive loss, the framework creates a shared continuous space where cross-domain structural motifs survive federation. After pre-training, class-wise tokens and prompt pools adapt the frozen encoders to node, edge, and graph classificati

What carries the argument

The load-bearing mechanism is the symmetric graph-text contrastive loss operating in a continuous embedding space, pairing graph encoder outputs with text summaries from a frozen pre-trained language model. Supporting mechanisms include: context-aware semantic summarization that concatenates node and one-hop neighborhood descriptions; decoupled structural-semantic encoding that separates domain-invariant topology from local semantics; history-matching that weights recent global structural encoders by similarity to local semantic anchors to stabilize updates; and topology-aware aggregation that weights clients by average degree. This machinery replaces codebook search and discrete token trans

Load-bearing premise

The paper's central claim rests on the assumption that continuous contrastive alignment—not some other component of the two-phase pipeline—is what produces the reported gains; the ablations test auxiliary modules but never remove the alignment itself.

What would settle it

Run FedGALA's full pipeline but swap the continuous alignment for a vector-quantized bottleneck while keeping history-matching, class-wise tokens, and prompt pools identical; if results remain close, the central mechanism is not continuous alignment. A lighter check is to ablate the contrastive loss itself and observe whether downstream performance falls to baseline levels.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A single pre-trained graph encoder can serve node, edge, and graph classification tasks through prompt adaptation alone, removing the need for task-specific federated training.
  • Per-round communication drops because only structural encoder parameters are uploaded, and client-side computation no longer scales with codebook or prompt-pool size.
  • Federated pre-training converges in roughly 6–11 rounds instead of 15–20, reducing the number of synchronization rounds in bandwidth-limited settings.
  • Few-shot and cold-start clients benefit: 2-shot accuracy improves by up to 21.54% on WikiCS over the best baseline, suggesting transferable knowledge survives federation.
  • Decoupling semantic and structural encoders prevents domain-specific features from being averaged away, preserving long-tail local knowledge that quantization tends to homogenize.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Extension: if continuous alignment is the operative cause, vector-quantized federated graph foundation models may be carrying avoidable complexity, and a simpler frozen-language-model plus trainable-structural-encoder design could become the default for bandwidth-limited federations.
  • Extension: because only the structural encoder is aggregated, the approach implicitly bets that textual semantics can stay local; a natural stress test is to increase the number of clients per dataset or introduce text-domain shift to see whether the global structural encoding remains sufficient.
  • Extension: the same two-phase scaffold—federated contrastive pre-training plus prompt-based fine-tuning—could extend to other paired modalities, such as graph-image or graph-tabular data, a direction the paper mentions only as future multi-modal work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. FedGALA is a two-phase federated graph foundation model (FedGFM) framework. In the pre-training phase, clients train a graph encoder whose trainable structural encoder is aligned with a frozen PLM through an unsupervised contrastive loss in a continuous embedding space, while a separate semantic encoder is kept local. A history-matching module aggregates the R most recent global structural encoders, and class-wise prototypes are aggregated into global tokens. In the fine-tuning phase, the backbones are frozen and clients select and update soft prompts and graph prompts from global pools, followed by group-aware prompt aggregation. The paper claims that replacing discrete vector-quantized backbones with continuous graph-text alignment (as in FedGFM+ and FedBook) avoids 'irreversible knowledge loss' and yields up to 14.37% relative improvements over 22 baselines, with faster convergence and lower communication cost.

Significance. If the central claim were fully supported, the paper would make a useful contribution by questioning the necessity of VQ in FedGFMs and proposing a concrete continuous-alignment alternative that combines contrastive pre-training, prompt tuning, and topology-aware federation. The framework is systematically described, the experimental scope is broad (8 datasets, 3 tasks, 22 baselines), and the reported gains over FedGFM+ and FedBook are substantial. However, the paper currently does not provide a controlled test of its headline mechanism: no experiment replaces continuous alignment with a VQ bottleneck inside the same two-phase pipeline, and no ablation removes or weakens the continuous-alignment module itself. As a result, the causal attribution to 'continuous vs. discrete' is unsupported. The paper also ships no code, data, or detailed hyperparameters, which limits reproducibility. These issues are load-bearing for the paper's central claim and require substantial revision before the results can be assessed.

major comments (5)
  1. [Abstract; §4.4, Table 3] The central claim that continuous graph-text alignment (rather than VQ) drives FedGALA's gains is not tested. Table 3 ablates only the history-matching module and the global prompt pool; the continuous-alignment mechanism is never removed or replaced. The text says the pre-training phase is 'independently validated by the 2-shot results in Table 2,' but Table 2 reports the full FedGALA pipeline (pre-training + prompt fine-tuning + class-wise tokens), so it cannot isolate the alignment mechanism. A controlled comparison is needed: e.g., FedGALA with a VQ bottleneck (or a continuous variant of FedBook/FedGFM+) under identical pre-training objective, encoder, aggregation, and prompt-tuning protocol. Additionally, the paper asserts 'irreversible knowledge loss' in VQ methods without reporting any quantization-error or information-retention metric.
  2. [§4.1; §4.2, Table 1] The comparison against FedGFM+ and FedBook confounds multiple factors simultaneously: backbone architecture (VQ codebook vs. GraphSAGE+Sentence-BERT), pre-training objective, aggregation scheme, and fine-tuning protocol. Even if FedGALA surpasses these baselines by 5.41% (node), 4.49% (edge), and 4.17% (graph) on average, those gains cannot be attributed to the continuous-vs-discrete choice. The authors should report an experiment that holds all pipeline components fixed except the tokenization/alignment mechanism, or explicitly qualify the performance claim as a whole-framework comparison rather than a mechanism attribution.
  3. [§4.1 and Appendix references] The experimental setup is insufficiently detailed for a state-of-the-art claim with up to 14.37% improvement. The main text defers to appendices, but the provided manuscript's appendices do not contain training hyperparameters (learning rate, batch size, number of local epochs, optimizer, prompt dimensions, fine-tuning rounds, history length R, temperature initialization), the exact data partitioning and label split details for each dataset, or the number of runs and seeds beyond '10 standardized runs.' No code or data are released. The paper should provide the full setup in the main text or a complete appendix, and ideally release code/data to allow verification.
  4. [§4.6, Fig. 3] The convergence claim is not precisely defined. The text states FedGALA stabilizes at round 6/9/11 vs. 15-20 for baselines, but no convergence criterion (e.g., validation accuracy threshold, loss plateau) is given, and the figure (as presented) does not show error bars or a quantitative definition. Please define the criterion and report a table or figure with standard deviations. Also, the communication-cost claim ('transmitting only structural encoder parameters') should account for the history pool, class-wise prototypes, and prompt pool uploads/downloads; the current claim in §D ('FedGALA optimizes communication to O(|Θ|)') appears inconsistent with these additional transmissions.
  5. [§3.1.1, Eq. (5)] The topology-aware aggregation weight based on average degree is introduced without any justification beyond a heuristic statement. No sensitivity analysis or comparison to equal weighting (α_k = 1/K) or other topology-aware schemes (e.g., FedGTA's moments) is provided. Since this choice affects the global structural encoder in every round, the paper should either demonstrate its necessity through an ablation or present it as a design choice with supporting evidence.
minor comments (6)
  1. [§2.1, Eq. (1)] The global contrastive loss L_pre in Eq. (1) is written as a weighted sum but the actual local loss used in Eq. (4) is an InfoNCE-style symmetric loss. Clarify how Eq. (1) and Eq. (4) relate, and specify how α_k is normalized.
  2. [§3.1.2, Eq. (8)] The similarity aggregation in Eq. (8) is ambiguous: the softmax input is written as a sum over i∈D_k of S_r^(i), but it is unclear whether this sum is over the full local dataset or a batch, and whether the softmax is over the R history entries. Please clarify the notation and the normalization.
  3. [§3.1.3, Eq. (10)] The 'knowledge clarity' term uses z_i^⊤ z_j / (∥z_i∥∥z_j∥) but the denominator norm notation is missing in the displayed equation; also define N_k(i) and the relation between z_i and z^G_i (the graph encoder output used in Eq. (11)).
  4. [Appendix D] The complexity analysis states FedGALA's client-side complexity is O(L(|V|d^2+|E|d)+|V|d^2), but the |V|d^2 term for the contrastive alignment is not derived. Also, the communication cost O(|Θ|) ignores the R-history broadcast and class-wise prototype uploads; please correct or clarify.
  5. [Table 1] The 'Linear' baseline in Table 1 is not described in Section B of the appendix. Please specify its input features and training protocol.
  6. [References] Several references are incomplete or missing years (e.g., 'Chen et al.' with no year, 'Huang et al.' with no year, 'GQT* (Wang et al.)' with no year). These should be completed for reproducibility and scholarly rigor.

Circularity Check

0 steps flagged

No circular derivation; the main weakness is an unablated central mechanism, an attribution gap rather than circularity.

full rationale

This is an empirical federated learning paper, not a derivation whose output is forced by its inputs. FedGALA combines known components—contrastive graph-text alignment, decoupled structural/semantic encoders, history-matching, and prompt pools—and its claims are supported by comparisons against 22 baselines, including external non-overlapping methods. I find no equation-level circularity: no metric is defined in terms of a fitted parameter and then reported as a prediction; no uniqueness theorem or prior result by the same authors is invoked to forbid alternatives; and the constructive equations (contrastive loss Eq. 4, topology-aware aggregation Eq. 6, history-matching Eqs. 7–9, prompt aggregation Eq. 14) estimate values from data rather than defining the target conclusion. The self-citations (FedTAD, FedGFM+, FedBook) are to prior baselines or critiqued predecessors and are not load-bearing: the claim that VQ methods 'suffer from irreversible knowledge loss' is an assertion accompanied by comparative experiments, not an imported theorem. The most serious weakness is experimental attribution, not circularity. Section 4.4 explicitly declines to ablate the pre-training phase because 'the pre-training phase’s efficacy is independently validated by the 2-shot results in Table 2,' yet Table 2 evaluates the full FedGALA pipeline, so it cannot isolate the continuous-alignment mechanism. Likewise, comparisons against FedGFM+/FedBook vary backbone, objective, and aggregation simultaneously, leaving the headline 14.37% gain under-attributed. That is an unsupported mechanism claim—a correctness and experimental-design concern—rather than a circular derivation. Score 1 reflects the presence of minor self-citation with no load-bearing circularity.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The approach rests on several untested domain assumptions (continuous alignment transfers structural knowledge in federated settings; degree-weighted aggregation; one-hop summaries suffice). It introduces prompt pools, class-wise tokens, and history-matching as internal mechanisms, but their behavior is only demonstrated through ablation-style experiments, not external verification.

free parameters (6)
  • history length R = 5 (Sec. 3.1.2)
    R=5 is chosen without sensitivity analysis or justification; the paper does not show how performance varies with R.
  • token length for soft prompts = 16
    Figure 2 shows a sensitivity sweep over token lengths; the recommended configuration is 16. This is a hyperparameter, not a scientific constant.
  • global prompt pool size M = 10
    Figure 2 sweep over pool sizes; 10 is adopted. The paper acknowledges that larger pools (15) under-perform due to under-updated prompts.
  • learnable temperature τ in Eq. 3 = learned
    τ is a learnable temperature in the contrastive loss; its fitted value is not stated.
  • federated rounds T, T_ft = not reported
    Number of pre-training and fine-tuning rounds is not reported anywhere in the main text or appendices; Section 4.6 mentions stability round numbers for convergence plots but not the maximal rounds.
  • trade-off weights in task loss Eq. 13 = not reported
    The exact composition of L_task (e.g., classification loss plus contrastive or alignment terms) is not specified.
axioms (6)
  • domain assumption Unsupervised contrastive alignment between GNN and frozen PLM embeddings yields transferable structural knowledge in federated settings.
    This is the central premise of the approach. It is assumed from the start and is never tested against a VQ-based pre-training baseline under identical fine-tuning conditions (Sec. 4.4).
  • domain assumption Only transmitting structural encoder parameters θ_str prevents semantic drift and preserves privacy.
    The paper assumes local semantic encoders remain safe and that structural parameters suffice for global knowledge (Sec. 3.1.1), but it does not test whether shared structural encoders leak topology-sensitive information.
  • domain assumption Text summaries of 1-hop neighborhoods are sufficient semantic anchors.
    Context-aware semantic summarization (Sec. 3.1.1 (1)) uses only target plus one-hop descriptions, limiting text encoder context; the paper does not compare to larger neighborhoods.
  • ad hoc to paper Weighting clients by average degree α_k improves structural representation learning.
    Eq. 5 weights clients by average degree. This is a design choice with no ablation or justification beyond intuition; it could bias aggregation toward dense clients.
  • ad hoc to paper Class-wise prototypes computed with knowledge strength and clarity improve prompt initialization.
    Eq. 10-11 introduce these metrics as useful, but no ablation isolates them from other components.
  • domain assumption GraphSAGE and Sentence-BERT (768-dim) are appropriate encoders for all tasks.
    Sec. 4.1 states the architecture; no comparison with other GNN backbones or PLMs is given.
invented entities (3)
  • Federated history-matching module with temporal log of R global encoders no independent evidence
    purpose: Mitigates client drift and catastrophic forgetting by blending historical global models with local semantic parameters using similarity weights β.
    This is a new algorithmic component proposed by the paper. It is only evaluated via an ablation (Table 3); there is no external validation or formal analysis.
  • Group-aware prompt aggregation with latent groups S_m and dual prompt pools P_T, P_G no independent evidence
    purpose: Preserves task specialization and prevents dilution of specialized prompts.
    The clustering/grouping procedure is only vaguely described; no formal properties or external checks are given.
  • Global class-wise tokens aggregated via clustering of local prototypes no independent evidence
    purpose: Provides semantic anchors for frozen PLMs in fine-tuning.
    The 'clustering-based aggregation strategy' is not specified in detail; it is an algorithmic invention without external evidence.

pith-pipeline@v1.3.0-alltime-deepseek · 18128 in / 8570 out tokens · 71470 ms · 2026-08-04T06:12:32.629328+00:00 · methodology

0 comments
read the original abstract

Recent studies of federated graph foundational models (FedGFMs) break the idealized and untenable assumption of having centralized data storage to train graph foundation models, and accommodate the reality of distributed, privacy-restricted data silos. Despite their simplicity and intuition, existing studies that project aligned generalizable knowledge onto a discrete token space via vector-quantized backbones suffer from irreversible knowledge loss during the quantization process. In this context, we argue that reconciling the semantic-structural orthogonality and integrity between pre-trained language models (PLMs) and graph neural networks (GNNs) is paramount for developing effective FedGFMs while simultaneously mitigating the severe data heterogeneity and communication constraints inherent in distributed, resource-limited environments. To address these issues, we propose FedGALA (Federated Graph And Language Alignment), a framework that resolves graph-based semantic-structural orthogonality and integrity in federated settings by employing unsupervised contrastive learning to align GNNs and frozen PLMs within a continuous embedding space, thereby capturing robust, transferable general knowledge. Subsequently, FedGALA leverages a communication-efficient prompt tuning mechanism to steer these pre-aligned encoders and frozen PLMs, facilitating effective adaptation to diverse downstream tasks while circumventing the prohibitive overhead of full-parameter fine-tuning. The comprehensive experiments validate that FedGALA outperforms all competitive baselines across multi-domain datasets on multiple tasks with up to 14.37% performance improvement.

Figures

Figures reproduced from arXiv: 2601.21369 by Di Wu, Guocong Quan, Miao Hu, Xianzhi Zhang, Xunkai Li, Yinlin Zhu, Yuming Ai.

Figure 1
Figure 1. Figure 1: FedGALA framework contains two training phases: ❶ Federated Pre-training Graph Encoder: aligning frozen PLMs with partitioned graph encoders (structural and semantic) via contrastive alignment and history-matching to derive global structural parameters, class-wise tokens, and local semantic encoders. ❷ Local Prompt-based Federated Fine-tuning: adapting the frozen PLMs to various downstream tasks, and the t… view at source ↗
Figure 3
Figure 3. Figure 3: Convergence rates during the pre-training phase, where the proposed FedGALA consistently achieves faster convergence than all baseline methods. 4.6. Efficiency Analysis (Answer for Q5) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Sensitivity analysis of FedGALA. preserving task specialization. Without this mechanism, FedGALA suffers drops of 4.82% on OGB-arxiv and 4.59% on FB15K237, underscoring the necessity of mitigating the knowledge entanglement resulting from varying localized knowledge for efficient downstream adaptation. 4.5. Sensitivity Analysis (Answer for Q4) We conduct a sensitivity analysis to study the impact of token-… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 7 linked inside Pith

  1. [5]

    Liu, H., Feng, J., Kong, L., Liang, N., Tao, D., Chen, Y ., and Zhang, M

    URL https://arxiv.org/abs/ 2410.14961. Liu, H., Feng, J., Kong, L., Liang, N., Tao, D., Chen, Y ., and Zhang, M. One for all: Towards training one graph model for all classification tasks.arXiv preprint arXiv:2310.00149,

    Pith/arXiv arXiv

  2. [8]

    Representing text for joint embedding of text and knowledge bases

    Toutanova, K., Chen, D., Pantel, P., Poon, H., Choudhury, P., and Gamon, M. Representing text for joint embedding of text and knowledge bases. InProceedings of the 2015 conference on empirical methods in natural language processing, pp. 1499–1509,

    2015

  3. [9]

    URL https://arxiv.org/abs/ 2510.07755. Xia, L. and Huang, C. Anygraph: Graph foundation model in the wild.arXiv preprint arXiv:2408.10700,

    arXiv

  4. [10]

    URL https://doi.org/ 10.1145/3394486.3403168

    1145/3394486.3403168. URL https://doi.org/ 10.1145/3394486.3403168. Tan, Y ., Liu, Y ., Long, G., Jiang, J., Lu, Q., and Zhang, C. Federated learning on non-iid graphs via structural knowl- edge sharing. InProceedings of the AAAI conference on artificial intelligence, volume 37, pp. 9953–9961,

    arXiv

  5. [11]

    ISBN 9798400714542

    Association for Comput- ing Machinery. ISBN 9798400714542. doi: 10.1145/ 3711896.3737410. URL https://doi.org/10. 1145/3711896.3737410. Yang, Z., Cohen, W. W., and Salakhutdinov, R. Revisit- ing semi-supervised learning with graph embeddings. In Proceedings of the 33rd International Conference on In- ternational Conference on Machine Learning, ICML, pp. 40–48,

    arXiv

  6. [12]

    Federated graph learning–a position paper.arXiv preprint arXiv:2105.11099, 2021a

    Zhang, H., Shen, T., Wu, F., Yin, M., Yang, H., and Wu, C. Federated graph learning–a position paper.arXiv preprint arXiv:2105.11099, 2021a. Zhang, K., Yang, C., Li, X., Sun, L., and Yiu, S. M. Sub- graph federated learning with missing neighbor genera- tion.Advances in Neural Information Processing Systems, NeurIPS, 2021b. Zhang, Z., Bu, J., Ester, M., Z...

    Pith/arXiv arXiv 1911

  7. [14]

    To- wards effective federated graph foundation model via mitigating knowledge entanglement.arXiv preprint arXiv:2505.12684, 2025a

    Zhu, Y ., Li, X., Jia, J., Hu, M., Wu, D., and Qiu, M. To- wards effective federated graph foundation model via mitigating knowledge entanglement.arXiv preprint arXiv:2505.12684, 2025a. 10 Rethinking Federated Graph Foundation Models: A Graph-Language Alignment-based Approach Zhu, Y ., Shi, H., Wang, X., Liu, Y ., Wang, Y ., Peng, B., Hong, C., and Tang, ...

    arXiv 2025

  8. [16]

    Due to its emphasis on preserving structural integrity, GIN is frequently the preferred choice for graph-level representation tasks

    is engineered to maximize structural discriminative power, theoretically matching the Weisfeiler- Lehman graph isomorphism test in its capacity to distinguish complex graph topologies. Due to its emphasis on preserving structural integrity, GIN is frequently the preferred choice for graph-level representation tasks. ❷ FGL Approaches.We benchmark our work ...

    2017

  9. [17]

    The specific characteristics of these baselines are outlined below: FedAvg(McMahan et al.,

    and FedStar (Tan et al., 2023)). The specific characteristics of these baselines are outlined below: FedAvg(McMahan et al.,

    2023

  10. [18]

    facilitates graph-level FGL by separating the learning of structures and features. Clients utilize an independent encoder to share domain-invariant structural information while maintaining personalized local encoders for features, thereby minimizing misalignment across the federation. ❸ Federated Adaptations of Centralized GFM Approaches.This category ada...

    2023

  11. [19]

    introduces a systematic server-side aggregation strategy designed to optimize the global codebook by balancing intra-domain coherence with inter-domain diversity. In the first phase, the server identifies semantically similar knowledge units across clients and utilizes a frequency-guided alignment mechanism where low-frequency (potentially unreliable) tok...

    2025

  12. [64]

    Regarding theFederated Adaptations of Centralized GFM Methods, andFedGFMswe strictly maintain the architectural backbones as reported in their respective original studies

    ForFL/FGL Methods, we employ task-specific backbones in instances where a custom architecture is not specified: GraphSAGE is utilized for node-level and edge-level classification, while GIN is adopted for graph-level tasks. Regarding theFederated Adaptations of Centralized GFM Methods, andFedGFMswe strictly maintain the architectural backbones as reported...

    2017

  13. [2017]

    and Cangea, C

    Mernyei, P. and Cangea, C. Wiki-cs: A wikipedia-based benchmark for graph neural networks.arXiv preprint arXiv:2007.02901,

    Pith/arXiv arXiv 2007

  14. [2019]

    Fedtad: Topology-aware data-free knowledge distilla- tion for subgraph federated learning.arXiv preprint arXiv:2404.14061,

    Zhu, Y ., Li, X., Wu, Z., Wu, D., Hu, M., and Li, R.-H. Fedtad: Topology-aware data-free knowledge distilla- tion for subgraph federated learning.arXiv preprint arXiv:2404.14061,

    Pith/arXiv arXiv

  15. [2020]

    ISBN 9781713829546

    Curran Associates Inc. ISBN 9781713829546. 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. InProceedings of the ACM on Web Conference 2025, pp. 1142–1153,

    2025

  16. [2021]

    P., Miao, S., Li, P., and Ying, R

    Chen, J., Zuo, H., Wang, H. P., Miao, S., Li, P., and Ying, R. Gfse: A foundational model for graph structural encod- ing. Chen, J., Li, C., Li, G., Hopcroft, J. E., and He, K. Rethink- ing tokenized graph transformers for node classification. arXiv preprint arXiv:2502.08101,

    Pith/arXiv arXiv

  17. [2023]

    Li, X., Wu, Z., Zhang, W., Sun, H., Li, R.-H., and Wang, G

    Curran As- sociates Inc. Li, X., Wu, Z., Zhang, W., Sun, H., Li, R.-H., and Wang, G. Adafgl: A new paradigm for federated node clas- sification with topology heterogeneity.arXiv preprint arXiv:2401.11750, 2024a. Li, X., Wu, Z., Zhang, W., Zhu, Y ., Li, R.-H., and Wang, G. Fedgta: Topology-aware averaging for federated graph learning.arXiv preprint arXiv:2...

    Pith/arXiv arXiv

  18. [2024]

    Li, Q., He, B., and Song, D

    URL https: //arxiv.org/abs/2407.09709. Li, Q., He, B., and Song, D. Model-contrastive federated learning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR, pp. 10713–10722,

    Pith/arXiv arXiv

  19. [2025]

    ISBN 9798400712456

    Asso- ciation for Computing Machinery. ISBN 9798400712456. doi: 10.1145/3690624.3709277. URL https://doi. org/10.1145/3690624.3709277. Hu, W., Fey, M., Zitnik, M., Dong, Y ., Ren, H., Liu, B., Catasta, M., and Leskovec, J. Open graph benchmark: Datasets for machine learning on graphs.Advances in neu- ral information processing systems, NeurIPS, 33:22118– 22133,

    arXiv