Pith. sign in

REVIEW 4 major objections 4 minor 60 references

FedGAMMA claims that a federated, privacy-preserving pretraining pipeline can unify image, text, and graph-topology learning on multimodal graphs, outperforming existing methods on every evaluated dataset.

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

2026-08-01 22:34 UTC pith:TOOQNYQE

load-bearing objection Plausible architecture for a real gap, with consistent reported gains, but the central empirical claim is not independently checkable and Theorem 2 is circular; worth refereeing if the authors open the box. the 4 major comments →

arxiv 2607.15687 v1 pith:TOOQNYQE submitted 2026-07-17 cs.LG

Toward Federated Multimodal Graph Foundation Models: A Topology-Aware Multimodal Alignment Framework

Xunkai Li , Guohao Fu , Yuming Ai , Zhengyu Wu , Hongchao Qin , Rong-Hua Li , Guoren Wang This is my paper
classification cs.LG
keywords federated learninggraph foundation modelsmultimodal graphsmodality alignmentdual-channel aggregationprompt tuningoptimal transportprivacy-preserving learning
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 proposes FedGAMMA, a two-stage framework for federated multimodal graph foundation models. It tries to establish that a single model can be pretrained across privacy-separated clients—each holding a graph whose nodes carry both image and text features—and then adapted to downstream tasks with lightweight prompts, without ever sharing raw data. The authors argue that three design choices are essential: disentangling shared cross-modal signal from modality-specific info via optimal transport, decoupling semantic and structural views through residual graphs and positional encodings, and aggregating client updates separately along feature and graph channels using centroid-based similarity. If correct, this would enable collaborative, transferable multimodal graph learning in regulated or fragmented settings where centralizing data is impossible.

Core claim

The central claim is that a federated multimodal graph foundation model can be built by reformulating the problem as two-stage semantic-structural alignment. In pretraining, a shared-private semantic enhancer separates cross-modal commonality from modality-specific detail and aligns the shared streams with a Sinkhorn–Wasserstein optimal-transport loss. A topology-aware fusion module propagates the fused features over a semantic residual graph and the original topology, while a separate structural-role view captures connectivity via dual positional encodings and a diffusion graph. The two views are combined only after independent formation. On the server, a dual-channel affinity aggregator co

What carries the argument

The load-bearing mechanism is the dual-channel affinity-aware aggregation combined with shared-private modality disentanglement. Clients upload compact K-means centroids per channel—one from the fused embeddings and one from structural-role embeddings—each paired with a spectral energy descriptor (normalized Laplacian eigenvalues of the cluster’s cosine-similarity graph). The server compares clients through MMD between centroid sets, producing two affinity matrices that route the multimodal-fusion and graph-fusion parameter groups separately. A theoretical result (Theorem 1) shows that this strictly dominates any single-channel aggregation scheme whenever feature-channel and graph-channel cl

Load-bearing premise

The headline result depends on the evaluation being fair and representative: the twelve datasets come from a benchmark built by the same group, the five-client Louvain/Metis partitions are assumed to capture real-world non-IID heterogeneity, and the nine baselines are author-implemented federated adaptations that may be under-tuned; if any of these fail, the claim of outperforming all baselines on all twelve tasks is not established.

What would settle it

A concrete refutation would be an independent replication on a held-out third-party multimodal graph benchmark (not from the same group) where a properly tuned implementation of the strongest baseline (e.g., Fed-PLANET or MM-FedGALA) matches or exceeds FedGAMMA’s accuracy, or where a formal privacy analysis shows that the uploaded centroids and Laplacian spectral descriptors can be inverted to recover identifiable node-level information.

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

If this is right

  • If FedGAMMA’s claims hold, organizations can collaboratively pretrain a multimodal graph foundation model across data silos (e.g., hospitals or retailers) without centralizing sensitive node attributes, while still obtaining a shared transferable encoder.
  • The reported 2–5× reduction in communication rounds to reach a given accuracy suggests that channel-wise aggregation accelerates convergence, which would lower the cost of federated training in practice.
  • The few-shot results imply that the pretrained representation retains high-quality cross-modal and structural geometry, so downstream tasks with very few labels can benefit without full fine-tuning.
  • The ablation identifying the cross-modal enhancer as the dominant component indicates that explicit shared-private separation—rather than simple feature concatenation—is necessary for federated multimodal graphs.
  • The architecture’s reliance on compact centroids rather than raw data provides a practical blueprint for privacy-preserving graph federation, one that could be extended to other modalities or to formal differential-privacy guarantees.

Where Pith is reading between the lines

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

  • If the dual-channel dominance theorem is robust, a general principle emerges: in federated learning with heterogeneous clients, any parameter group that serves a distinct inductive role should be aggregated using its own similarity metric; this may apply beyond graphs to other multi-view or multi-objective federated settings.
  • The spectral energy descriptor in the centroids suggests that sharing information about the internal geometry of clusters (via Laplacian spectra) can improve similarity estimation; one testable extension is whether replacing K-means with other clustering or sketching methods preserves the gains while offering stronger privacy.
  • The prompt-pool design with UCB-style exploration implies that federated model adaptation can be treated as a multi-armed bandit over shared knowledge; a direct test would be whether the pool size becomes more important as client heterogeneity grows beyond the five-client Louvain partitions studied here.

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

4 major / 4 minor

Summary. The paper proposes FedGAMMA, a two-stage federated multimodal graph foundation model for multimodal-attributed graphs (MAGs) whose nodes carry image and text features from frozen CLIP encoders. It targets three challenges: privacy-constrained cross-modal alignment, semantic-structural grounding, and heterogeneity-aware aggregation. Pre-training uses a shared-private semantic enhancer with optimal-transport alignment, a topology-aware graph fusion module that decouples semantic and structural views via residual/diffusion graphs and dual positional encodings, and a dual-channel MMD-based affinity aggregation mechanism. Fine-tuning uses graph-aware prompts, a shared prompt pool with UCB-style selection, and channel-wise federated synchronization. Experiments on twelve MM-OpenFGL datasets report the best result on all twelve cells across six downstream tasks, with gains up to 12.96%, plus few-shot results and efficiency/convergence analyses. Three theorems are claimed: dual-channel dominance, local pre-training bounds on cross-modal/structural discrepancy, and nonconvex convergence.

Significance. If the empirical results are reproducible, FedGAMMA is a valuable contribution at the underexplored intersection of federated learning, multimodal learning, and graph foundation models. Strengths include 10-seed mean±std reporting, ablations that isolate the cross-modal enhancer as the dominant component, a credible efficiency analysis, and an honest limitations section that acknowledges missing formal privacy analysis. The main risk is that the central SOTA claim is not independently checkable: the benchmark is from the same group, baselines are author-implemented federated adaptations, and no code or full hyperparameter configurations are released. The theoretical support is currently not load-bearing, and Theorem 1 contains a mathematical error. With code release, transparency about baseline tuning, significance testing, and repair or de-emphasis of the theory, the paper could meet the journal's bar.

major comments (4)
  1. [§VI (Theorem 1, Eqs. (12)–(14))] The theorem's quantitative claim is inconsistent. δ_i is defined in Eq. (12) as ∥w_f − w_g∥_2 (a norm), but Eq. (14) gives a gap of μ/2 δ_i, and the proof sketch derives μ/4 ∥w_f − w_g∥^2 = μ/4 δ_i^2. Strong convexity yields a squared-norm penalty, so the correct gap is μ/4 δ_i^2 − 2ε and the positivity condition is δ_i^2 > 8ε/μ, not δ_i > 4ε/μ. This does not destroy the qualitative dominance claim but the stated bound and threshold are wrong.
  2. [§VI (Theorem 2, Eq. (15); Assumptions A1–A4)] Theorem 2 is circular: it assumes L_align ≥ μ1 Δ_mm and L_nce + L_reg ≥ μ2 Δ_ss, then substitutes these lower bounds into F_k to conclude Δ_mm + Δ_ss ≤ F_k/μ1 + F_k/μ2. The conclusion is the assumed relation rearranged; no argument connects the losses to the discrepancies. Theorems 1 and 3 also invoke Assumptions A1–A4, which are never stated anywhere in the manuscript. Please state the assumptions and prove a genuine relation, or present the theory as motivational rather than as an explanation of observed behavior.
  3. [§VII-A, §VII-B, Table II] The central empirical claim—'best result on all twelve cells' with up to 12.96% gains—is not independently verifiable. All twelve datasets come from MM-OpenFGL [49], developed by the same group; all nine baselines are author-implemented federated adaptations; and the paper does not release code, data splits, seeds, or per-baseline hyperparameter search ranges. 'Tuned on a held-out validation split' (§VII-A) is insufficient to rule out asymmetric tuning. In addition, no significance tests accompany Table II; on Toys generation the margin over MM-FedGFM is 19.42±0.45 vs 19.07±0.51, within one standard deviation. Please provide the artifact release, per-baseline tuning budgets, and paired significance tests.
  4. [§V-D (Eq. (5)) and §VIII] The privacy motivation is not backed by a formal privacy analysis. K-means centroids and normalized Laplacian spectral descriptors are shared with the server, and the paper claims this 'does not expose raw data,' but no differential privacy, information-theoretic leakage, or empirical attack analysis is provided; §VIII concedes that a formal privacy analysis 'remains valuable.' Since privacy preservation is a stated motivation (C1, C3), the privacy claim should be either formally analyzed or explicitly qualified as only 'raw data are not transmitted' with a discussion of residual inference risk.
minor comments (4)
  1. [Throughout] Typos: 'accross' (abstract), 'seperated' (§II-B), 'adde' (Algorithm 1).
  2. [Fig. 3] The axis labels in Fig. 3 are rendered as garbled /uni0000... entities; the figure must be regenerated to be readable.
  3. [§IV] The three 'observations' are presented as an empirical investigation, but no quantitative protocol, metrics, or measurements are given; Fig. 1 is a schematic. Please either add a small quantitative study or label these as motivating illustrations.
  4. [§VII-F] The '2-5× fewer communication rounds' claim is supported by Fig. 4, but the figure panels are not tied to named datasets, and Table V only covers Movies. Please specify the tasks/datasets in Fig. 4 and state which cells support the 2-5× claim.

Circularity Check

1 steps flagged

The central SOTA claim is an external benchmark comparison, not a derived prediction, but Theorem 2's 'bound' is a conditional tautology that assumes the very lower bounds it purports to establish.

specific steps
  1. other [§VI, Theorem 2 (Eq. 15) and proof sketch]
    "Suppose there exist μ1, μ2 > 0 such that the alignment loss satisfies L_align ≥ μ1 Δ_mm and the contrastive and regularization losses satisfy L_nce + L_reg ≥ μ2 Δ_ss. Then Δ_mm + Δ_ss ≤ 1/μ1 F_k(Θ) + 1/μ2 F_k(Θ). ... Substituting into the assumed lower bounds ... yields Δ_mm ≤ F_k(Θ)/μ1 and Δ_ss ≤ F_k(Θ)/μ2."

    The theorem's conclusion is obtained by substituting trivially true upper bounds (L_align ≤ F, L_nce + L_reg ≤ F) into its own assumed lower bounds. Thus Δ_mm ≤ F/μ1 is just the assumed L_align ≥ μ1 Δ_mm divided by μ1 and then weakened via L_align ≤ F; no architectural or data-dependent argument establishes the μ-lower bounds. The 'result' is therefore the assumption restated, not a derived prediction from the FedGAMMA architecture. The surrounding text ('Theorem 2 shows the local pre-training objective bounds ... discrepancy') presents this conditional tautology as an independent grounding of §IV.

full rationale

The main empirical claim — best results on all twelve MM-OpenFGL benchmark cells with gains up to 12.96% — is an experimental measurement against nine baselines; it is not obtained by fitting a parameter and then renaming it a prediction, and it does not reduce to a definition. The evaluation is difficult to audit (datasets from the same group's benchmark [49], author-implemented baselines, unreported hyperparameter ranges), but those are reproducibility/representativeness risks, not circular reductions. The one genuine circularity is narrow: Theorem 2 assumes μ-lower bounds connecting losses to discrepancies and then 'derives' the discrepancy bound from those assumptions, so the theoretical support is vacuous unless the lower bounds are proved elsewhere, which the paper does not do. §VIII openly concedes the absence of formal privacy analysis, a limitation rather than a circular step. Overall, the central contribution retains independent empirical content, so the score is moderate (4), reflecting the Theorem 2 tautology rather than global circularity.

Axiom & Free-Parameter Ledger

10 free parameters · 7 axioms · 0 invented entities

The ledger shows the paper's load-bearing external inputs: standard OT/MMD/contrastive machinery, frozen CLIP features, a simulated federated benchmark, and several unstated or merely assumed conditions (A1–A4; μ-lower-bounds; privacy of centroid sharing). The empirical SOTA claim depends mostly on datasets, baselines, and hyperparameters; the theoretical claims depend on unstated assumptions. None of these are fatal by themselves, but together they substantially reduce the amount of independent support the reader gets from the manuscript.

free parameters (10)
  • θ (Eq. 3) = not reported
    Coefficient balancing L_align and the disentanglement/reconstruction regularizer; chosen by validation tuning.
  • β and λ_reg (Eq. 4) = not reported
    Weights for contrastive loss and composite graph regularizer; chosen by validation tuning.
  • η (MMD bandwidth, Eq. 6) = not reported
    Bandwidth in the affinity kernel; controls how sharply client similarities become aggregation weights.
  • κ (UCB exploration, Eq. 11) = not reported
    Exploration coefficient in prompt-pool selection; affects pool diversity and fine-tuning behavior.
  • γ (prompt mixing weight) = not reported
    Residual weight fusing shared pool prompt with local graph-to-prompt prompt.
  • M (centroids per channel) = 5
    Number of K-means centroids used to summarize each client in each channel; set by hand.
  • d_c (spectral descriptor dimension) = 768 (default)
    Number of Laplacian eigenvalues in centroid descriptors; sensitivity test shows large performance swings (up to ~40%) with d_c.
  • K_pool and k_sel = 10 and 5
    Global prompt-pool size and number of selected prompts per client per round.
  • k_nn in semantic residual graph = not specified
    Number of nearest neighbors added to G_res; central to the semantic view but no value or sensitivity given.
  • diffusion graph parameters = not specified
    Diffusion hops and geometric decay for G_diff; structural-view construction but unspecified.
axioms (7)
  • standard math Sinkhorn optimal transport and MMD provide valid alignment and similarity measures for high-dimensional features.
    Invoked in §V-B and §V-D; standard but assumes enough samples and proper kernel bandwidth.
  • standard math InfoNCE/contrastive objectives align the structural and semantic views.
    Used in Eq. (4); assumes negative sampling and augmentation choices are adequate.
  • domain assumption Frozen CLIP image/text features are sufficient node modality representations.
    §II-A/§VII-A: all methods use CLIP d=768; if CLIP features are poor for a domain, the whole pipeline inherits that.
  • domain assumption The 5-client Louvain/Metis partitions simulate realistic non-IID federated heterogeneity.
    §VII-A: the main results depend on this simulation, but no real federated deployment is tested.
  • ad hoc to paper Assumptions A1–A4 (smoothness, bounded variance, etc.) hold for Theorems 1 and 3.
    §VI references A1–A4 but never defines them; the convergence and dominance guarantees are conditional on unspecified conditions.
  • ad hoc to paper There exist μ1, μ2>0 such that L_align ≥ μ1 Δ_mm and L_nce + L_reg ≥ μ2 Δ_ss in Theorem 2.
    §VI: the theorem's conclusion is the assumed lower bound divided by the objective; existence of such constants is asserted, not proved.
  • domain assumption Sharing K-means centroids and Laplacian spectral descriptors does not leak private raw data.
    §V-D/Algorithm 1: the privacy claim rests on this; no formal privacy analysis is provided and §VIII concedes this.

pith-pipeline@v1.3.0-alltime-deepseek · 25712 in / 19112 out tokens · 152881 ms · 2026-08-01T22:34:36.769534+00:00 · methodology

0 comments
read the original abstract

Multimodal-attributed graphs (MAGs), whose nodes carry modalities such as images and text alongside topological structure, now pervade applications including social platforms, e-commerce, and biomedical networks, offering richer semantic signals than single-modality graphs. In practice, such graphs are fragmented across privacy-restricted silos owned by different platforms and institutions, so learning a broadly transferable model over them demands collaborative training that never exposes raw data. This places the task at the intersection of multimodal graph learning and federated learning, yet existing methods cover only one side of it. To address the challenges from these two perspectives, we propose FedGAMMA, casting federated multimodal graph foundation learning as a two-stage semantic-structural alignment problem of federated pre-training and prompt-based fine-tuning. During pre-training, a shared-private semantic enhancer disentangles cross-modal commonality from modality-specific information, aligning it through optimal transport, a topology-aware graph fusion module decouples semantic and structural views via semantic residual graphs and dual positional encodings, and a dual-channel affinity-aware aggregation mechanism estimates client similarity from feature and graph centroids without exposing raw data. During fine-tuning, FedGAMMA adapts the pretrained encoder through lightweight graph-aware prompts, a shared prompt pool with controlled exploration, and channel-wise prompt synchronization. Experiments on twelve multimodal graph datasets show FedGAMMA consistently surpassing a broad range of baselines across downstream tasks, with gains of up to 12.96%. FedGAMMA further outperforms competitive baselines accross multi-domain datasets on multiple tasks with up to 5.71% under few-shot learning scenario.

Figures

Figures reproduced from arXiv: 2607.15687 by Guohao Fu, Guoren Wang, Hongchao Qin, Rong-Hua Li, Xunkai Li, Yuming Ai, Zhengyu Wu.

Figure 1
Figure 1. Figure 1: Comparison of FedGFM and FedMGFM for federated multi-modal graph learning. FedGFM suffers from modality collapse and misallocation due to single-weight aggregation, while FedMGFM adopts decoupled multi-view modeling and channel-wise aggregation to preserve topology. topological structure. FedMGFM addresses this by decoupling the encoding into complementary views: a semantic view that propagates over the or… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the FedGAMMA framework. The framework follows a two-stage pipeline. Pre-training (left): the Cross-Modal Semantic Enhancer disentangles shared and modality-private streams, and the Topology-Aware Graph Fusion module propagates features over complementary semantic and structural views. Federated synchronization (center): clients upload feature and graph centroids; the server computes dual-channe… view at source ↗
Figure 3
Figure 3. Figure 3: Sensitivity analysis of FedGAMMA on five MAG tasks. The top row sweeps the global prompt-pool size K, and the bottom row sweeps the modality-centroid dimensionality dc. Whiskers denote ±1 std, and the red star marks the default configuration. 5 10 15 20 25 30 35 40 45 50 Communication Round R 20 30 40 50 Accuracy (%) Node Cls. (Movies) 5 10 15 20 25 30 35 40 45 50 Communication Round R 50 55 60 65 70 75 AU… view at source ↗
Figure 4
Figure 4. Figure 4: Convergence analysis of pre-training. Each panel reports the validation metric of one task at R = 5, 10, . . . , 50 communication rounds. on the cross-modal tasks, since K moves Grocery retrieval by 3.4% and QB alignment by 2.5% but KU matching by only 0.1%, mirroring the Q3 finding that the prompt pool mainly refines cross-client structural priors. The dc profile is more fundamental and far more task-depe… 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

60 extracted references · 8 linked inside Pith

  1. [1]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clarket al., “Learning transferable visual models from natural language supervision,” inInternational conference on machine learning. PmLR, 2021, pp. 8748–8763

    2021

  2. [2]

    One for all: Towards training one graph model for all classification tasks,

    H. Liu, J. Feng, L. Kong, N. Liang, D. Tao, Y . Chen, and M. Zhang, “One for all: Towards training one graph model for all classification tasks,” inInternational Conference on Learning Representations (ICLR), 2024

    2024

  3. [3]

    Unigraph: Learning a unified cross- domain foundation model for text-attributed graphs,

    Y . He, Y . Sui, X. He, and B. Hooi, “Unigraph: Learning a unified cross- domain foundation model for text-attributed graphs,” inACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2025, pp. 448–459

    2025

  4. [4]

    Anygraph: Graph foundation model in the wild,

    L. Xia and C. Huang, “Anygraph: Graph foundation model in the wild,” arXiv preprint arXiv:2408.10700, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Advances and open problems in federated learning,

    P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummingset al., “Advances and open problems in federated learning,”Foundations and Trends in Machine Learning, vol. 14, no. 1–2, pp. 1–210, 2021

    2021

  6. [6]

    Federated learning for generalization, robustness, fairness: A survey and benchmark,

    W. Huang, M. Ye, Z. Shi, G. Wan, H. Li, B. Du, and Q. Yang, “Federated learning for generalization, robustness, fairness: A survey and benchmark,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 9387–9406, 2024

    2024

  7. [7]

    Subgraph federated learning with missing neighbor generation,

    K. Zhang, C. Yang, X. Li, L. Sun, and S. M. Yiu, “Subgraph federated learning with missing neighbor generation,” inAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  8. [8]

    Personalized subgraph federated learning,

    J. Baek, W. Jeong, J. Jin, J. Yoon, and S. J. Hwang, “Personalized subgraph federated learning,” inInternational Conference on Machine Learning (ICML), 2023

    2023

  9. [9]

    Towards effective fed- erated graph foundation model via mitigating knowledge entanglement,

    Y . Zhu, X. Li, J. Jia, M. Hu, D. Wu, and M. Qiu, “Towards effective fed- erated graph foundation model via mitigating knowledge entanglement,” arXiv preprint arXiv:2505.12684, 2025

    arXiv 2025

  10. [10]

    Fedbook: A unified federated graph foundation codebook with intra- domain and inter-domain knowledge modeling,

    Z. Wu, Y . Zhu, X. Li, Z. Qiu, R.-H. Li, G. Wang, and C. Zhou, “Fedbook: A unified federated graph foundation codebook with intra- domain and inter-domain knowledge modeling,” 2025

    2025

  11. [11]

    Sinkhorn distances: Lightspeed computation of optimal transport,

    M. Cuturi, “Sinkhorn distances: Lightspeed computation of optimal transport,” inAdvances in Neural Information Processing Systems (NeurIPS), 2013, pp. 2292–2300

    2013

  12. [12]

    Optimal transport for domain adaptation,

    N. Courty, R. Flamary, D. Tuia, and A. Rakotomamonjy, “Optimal transport for domain adaptation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 39, no. 9, pp. 1853–1865, 2017

    2017

  13. [13]

    Multimodal deep learning,

    J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y . Ng, “Multimodal deep learning,” inInternational Conference on Machine Learning (ICML), 2011

    2011

  14. [14]

    Mmgcn: Multi-modal graph convolution network for personalized recommenda- tion of micro-video,

    Y . Wei, X. Wang, L. Nie, X. He, R. Hong, and T.-S. Chua, “Mmgcn: Multi-modal graph convolution network for personalized recommenda- tion of micro-video,” inACM International Conference on Multimedia (MM), 2019, pp. 1437–1445

    2019

  15. [15]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” inInternational Conference on Learning Rep- resentations (ICLR), 2017

    2017

  16. [16]

    Communication-efficient learning of deep networks from decentralized data,

    B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y. Arcas, “Communication-efficient learning of deep networks from decentralized data,” inInternational Conference on Artificial Intelligence and Statistics (AISTATS), 2017

    2017

  17. [17]

    The power of scale for parameter-efficient prompt tuning,

    B. Lester, R. Al-Rfou, and N. Constant, “The power of scale for parameter-efficient prompt tuning,”arXiv preprint arXiv:2104.08691, 2021

    Pith/arXiv arXiv 2021

  18. [18]

    Learning to prompt for vision- language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision- language models,”International Journal of Computer Vision, vol. 130, pp. 2337–2348, 2022

    2022

  19. [19]

    A survey on self-supervised learning: Algorithms, applications, and future trends,

    J. Gui, T. Chen, J. Zhang, Q. Cao, Z. Sun, H. Luo, and D. Tao, “A survey on self-supervised learning: Algorithms, applications, and future trends,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 9052–9071, 2024

    2024

  20. [20]

    A bayesian federated learning framework with online laplace approxima- tion,

    L. Liu, X. Jiang, F. Zheng, H. Chen, G.-J. Qi, H. Huang, and L. Shao, “A bayesian federated learning framework with online laplace approxima- tion,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 1, pp. 1–16, 2024

    2024

  21. [21]

    Rethinking federated graph foundation models: A graph-language alignment-based approach,

    Y . Zhu, D. Wu, X. Zhang, Y . Ai, X. Li, M. Hu, and G. Quan, “Rethinking federated graph foundation models: A graph-language alignment-based approach,”arXiv preprint arXiv:2601.21369, 2026

    arXiv 2026

  22. [22]

    Cross-modal federated human activity recognition,

    X. Yang, B. Xiong, Y . Huang, and C. Xu, “Cross-modal federated human activity recognition,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 8, pp. 5345–5361, 2024

    2024

  23. [23]

    Federated feature augmentation and alignment,

    T. Zhou, Y . Yuan, B. Wang, and E. Konukoglu, “Federated feature augmentation and alignment,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 11 119–11 135, 2024

    2024

  24. [24]

    Fedgta: Topology-aware averaging for federated graph learning,

    X. Li, Z. Wu, W. Zhang, Y . Zhu, R.-H. Li, and G. Wang, “Fedgta: Topology-aware averaging for federated graph learning,” inProceedings of the VLDB Endowment, vol. 17, no. 1, 2024, pp. 41–50

    2024

  25. [25]

    Federated graph semantic and structural learning,

    W. Huang, G. Wan, M. Ye, and B. Du, “Federated graph semantic and structural learning,” inInternational Joint Conference on Artificial Intelligence (IJCAI), 2023, pp. 3830–3838

    2023

  26. [26]

    Fedtad: Topology- aware data-free knowledge distillation for subgraph federated learning,

    Y . Zhu, X. Li, Z. Wu, D. Wu, M. Hu, and R.-H. Li, “Fedtad: Topology- aware data-free knowledge distillation for subgraph federated learning,” inInternational Joint Conference on Artificial Intelligence (IJCAI), 2024, pp. 5716–5724

    2024

  27. [27]

    Federated graph classification over non-iid graphs,

    H. Xie, J. Ma, L. Xiong, and C. Yang, “Federated graph classification over non-iid graphs,” inAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  28. [28]

    Fedproto: Federated prototype learning across heterogeneous clients,

    Y . Tan, G. Long, L. Liu, T. Zhou, Q. Lu, J. Jiang, and C. Zhang, “Fedproto: Federated prototype learning across heterogeneous clients,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 36, no. 8, 2022, pp. 8432–8440

    2022

  29. [29]

    Adafgl: A new paradigm for federated node classification with topology heterogeneity,

    X. Li, Z. Wu, W. Zhang, H. Sun, R.-H. Li, and G. Wang, “Adafgl: A new paradigm for federated node classification with topology heterogeneity,” inIEEE International Conference on Data Engineering (ICDE), 2024

    2024

  30. [30]

    Federatedscope-gnn: Towards a unified, comprehensive and efficient package for federated graph learning,

    Z. Wang, W. Kuang, Y . Xie, L. Yao, Y . Li, B. Ding, and J. Zhou, “Federatedscope-gnn: Towards a unified, comprehensive and efficient package for federated graph learning,” inACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD), 2022

    2022

  31. [31]

    Utility- preserving federated graph learning with dual-perspective fairness,

    R. Luo, H. Huang, S. Yu, F. Yu, F. Xia, S. K. Das, and C. Zhang, “Utility- preserving federated graph learning with dual-perspective fairness,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2026

    2026

  32. [32]

    Federated learning of generalized linear causal networks,

    Q. Ye, A. A. Amini, and Q. Zhou, “Federated learning of generalized linear causal networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 10, pp. 6623–6636, 2024. IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. XX, NO. X, MONTH YEAR 14

    2024

  33. [33]

    Gft: Graph foundation model with transferable tree vocabulary,

    Z. Wang, Z. Zhang, N. Chawla, C. Zhang, and Y . Ye, “Gft: Graph foundation model with transferable tree vocabulary,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  34. [34]

    Variational graph auto-encoders,

    T. N. Kipf and M. Welling, “Variational graph auto-encoders,” in NeurIPS Workshop on Bayesian Deep Learning, 2016

    2016

  35. [35]

    Deep graph infomax,

    P. Veli ˇckovi´c, W. Fedus, W. L. Hamilton, P. Li `o, Y . Bengio, and R. D. Hjelm, “Deep graph infomax,” inInternational Conference on Learning Representations (ICLR), 2019

    2019

  36. [36]

    Multimodal learning with transform- ers: A survey,

    P. Xu, X. Zhu, and D. A. Clifton, “Multimodal learning with transform- ers: A survey,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 10, pp. 12 113–12 132, 2023

    2023

  37. [37]

    Campa: Efficient and aligned multimodal graph learning via decoupled propagation and aggregation,

    C. Wan, X. Li, Y . Zuo, H. Deng, S. Li, B. Fan, H. Qin, R.-H. Li, and G. Wang, “Campa: Efficient and aligned multimodal graph learning via decoupled propagation and aggregation,”arXiv preprint arXiv:2605.11468, 2026

    Pith/arXiv arXiv 2026

  38. [38]

    Tmte: Effective multimodal graph learning with task-aware modality and topology co-evolution,

    Y . Zhuet al., “Tmte: Effective multimodal graph learning with task-aware modality and topology co-evolution,”arXiv preprint arXiv:2603.27723, 2026

    arXiv 2026

  39. [39]

    Rolemag: Learning neighbor roles in multimodal graphs,

    Y . Zuo, X. Li, Z. Zhang, R.-H. Li, and G. Wang, “Rolemag: Learning neighbor roles in multimodal graphs,”arXiv preprint arXiv:2604.12271, 2026

    Pith/arXiv arXiv 2026

  40. [40]

    Ntsformer: A self-teaching graph transformer for multimodal isolated cold-start node classification,

    J. Hu, Y . He, Y . Li, B. Hooi, and B. He, “Ntsformer: A self-teaching graph transformer for multimodal isolated cold-start node classification,” inProceedings of the AAAI Conference on Artificial Intelligence (AAAI), 2026

    2026

  41. [41]

    Modality-independent graph neural networks with global transformers for multimodal recommendation,

    J. Hu, B. Hooi, B. He, and Y . Wei, “Modality-independent graph neural networks with global transformers for multimodal recommendation,” in Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), vol. 39, no. 11, 2025, pp. 11 790–11 798

    2025

  42. [42]

    Mario: Multimodal graph reasoning with large language models,

    Y . Sunet al., “Mario: Multimodal graph reasoning with large language models,”arXiv preprint arXiv:2603.05181, 2026

    arXiv 2026

  43. [43]

    Toward effective multimodal graph foundation model: A divide-and- conquer based approach,

    S. Liu, X. Li, D. Su, R. Zhang, H. Qin, R.-H. Li, and G. Wang, “Toward effective multimodal graph foundation model: A divide-and- conquer based approach,”arXiv preprint arXiv:2602.04116, 2026

    arXiv 2026

  44. [44]

    Unigraph2: Learning a unified embedding space to bind multimodal graphs,

    Y . He, Y . Sui, X. He, Y . Liu, Y . Sun, and B. Hooi, “Unigraph2: Learning a unified embedding space to bind multimodal graphs,” inACM Web Conference (WWW), 2025, pp. 1759–1770

    2025

  45. [45]

    Representation learning with contrastive predictive coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representation learning with contrastive predictive coding,”arXiv preprint arXiv:1807.03748, 2018

    Pith/arXiv arXiv 2018

  46. [46]

    Deep graph contrastive representation learning,

    Y . Zhu, Y . Xu, F. Yu, Q. Liu, S. Wu, and L. Wang, “Deep graph contrastive representation learning,”arXiv preprint arXiv:2006.04131, 2020

    Pith/arXiv arXiv 2006

  47. [47]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li `o, and Y . Bengio, “Graph attention networks,” inInternational Conference on Learning Representations (ICLR), 2018

    2018

  48. [48]

    A kernel two-sample test,

    A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Sch ¨olkopf, and A. Smola, “A kernel two-sample test,”Journal of Machine Learning Research, vol. 13, pp. 723–773, 2012

    2012

  49. [49]

    Mm-openfgl: A comprehensive benchmark for multimodal federated graph learning,

    X. Liet al., “Mm-openfgl: A comprehensive benchmark for multimodal federated graph learning,”arXiv preprint arXiv:2601.22416, 2026

    arXiv 2026

  50. [50]

    Fast unfolding of communities in large networks,

    V . D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast unfolding of communities in large networks,”Journal of Statistical Mechanics: Theory and Experiment, vol. 2008, no. 10, p. P10008, 2008

    2008

  51. [51]

    A fast and high quality multilevel scheme for partitioning irregular graphs,

    G. Karypis and V . Kumar, “A fast and high quality multilevel scheme for partitioning irregular graphs,”SIAM Journal on Scientific Computing, vol. 20, no. 1, pp. 359–392, 1998

    1998

  52. [52]

    Multiplex graph networks for multimodal brain network analysis,

    Z. Kong, L. Sun, H. Peng, L. Zhan, Y . Chen, and L. He, “Multiplex graph networks for multimodal brain network analysis,”ArXiv, vol. abs/2108.00158, 2021. [Online]. Available: https://api.semanticscholar. org/CorpusID:236772827

    Pith/arXiv arXiv 2021

  53. [53]

    Mgat: Multimodal graph attention network for recommendation,

    Z. Tao, Y . Wei, X. Wang, X. He, X. Huang, and T.-S. Chua, “Mgat: Multimodal graph attention network for recommendation,”Information Processing & Management, vol. 57, no. 5, p. 102277, 2020

    2020

  54. [54]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,”Journal of Machine Learning Research, vol. 21, no. 140, pp. 1–67, 2020

    2020

  55. [55]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” inConference on Empirical Methods in Natural Language Processing (EMNLP), 2019

    2019

  56. [56]

    Graphmae: Self-supervised masked graph autoencoders,

    Z. Hou, X. Liu, Y . Cen, Y . Dong, H. Yang, C. Wang, and J. Tang, “Graphmae: Self-supervised masked graph autoencoders,”arXiv preprint arXiv:2205.10803, 2022

    Pith/arXiv arXiv 2022

  57. [57]

    Graph con- trastive learning with augmentations,

    Y . You, T. Chen, Y . Sui, T. Chen, Z. Wang, and Y . Shen, “Graph con- trastive learning with augmentations,” inAdvances in Neural Information Processing Systems (NeurIPS), 2020

    2020

  58. [58]

    Neural discrete representation learning,

    A. van den Oord, O. Vinyals, and K. Kavukcuoglu, “Neural discrete representation learning,”Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  59. [59]

    Federated optimization in heterogeneous networks,

    T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,” inMachine Learn- ing and Systems (MLSys), 2020

    2020

  60. [60]

    Ditto: Fair and robust federated learning through personalization,

    T. Li, S. Hu, A. Beirami, and V . Smith, “Ditto: Fair and robust federated learning through personalization,”International Conference on Machine Learning (ICML), 2021. Xunkai Liis currently pursuing the PhD degree in Beijing Institute of Technology, advised by Prof. Rong-Hua Li. He received the BS degree from Shandong University in 2022. His research inte...

    2021