Pith. sign in

REVIEW 4 major objections 4 minor 127 references

FedTCR claims that the three-way heterogeneity of federated multimodal graph learning—across tasks, modalities, and graph topologies—can be handled by a two-stage, task-agnostic pre-training scheme with topology-aware cross-modal prototype

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

2026-08-05 00:30 UTC pith:72TFOUH4

load-bearing objection FedTCR is a coherent, well-tested method for federated multimodal graph learning; the routing mechanism is plausible but its distinctive selection step is not directly validated, and the heterogeneous-task evaluation needs external baselines. the 4 major comments →

arxiv 2608.00623 v1 pith:72TFOUH4 submitted 2026-08-01 cs.LG

Towards Effective Federated Multimodal Graph Learning via Navigating Multifaceted Heterogeneity

Yinlin Zhu , Di Wu , Yi Zhang , Xunkai Li , Wang Luo , Wei-Jin Huang , Miao Hu , Guocong Quan This is my paper
classification cs.LG
keywords federated learningmultimodal graph learningheterogeneityprototype routingcontrastive learninggraph neural networksdifferential privacytask-agnostic pre-training
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.

Federated multimodal graph learning lets organizations with privately held, modality-rich graphs train together without sharing raw data, but existing federated graph methods assume one task, one modality, and similar topology. This paper argues that in real decentralized graphs these assumptions fail in three ways at once: clients pursue different objectives, modalities differ in quality and semantic domain, and local graph structures diverge. FedTCR is proposed as the first systematic algorithm built for this setting: it separates task-agnostic federated pre-training from isolated task-oriented fine-tuning, and during pre-training it distills each client's modality knowledge into compact, topology-weighted prototypes, then has the server route each client's most similar and least similar other-client prototypes as contrastive references. The result, the paper claims, is a shared multimodal graph encoder that improves on existing methods for node classification, link prediction, cross-modal retrieval, and graph-conditioned text and image generation across seven domains. A sympathetic reader would care because, if correct, this is a path to collective intelligence over private multimodal graphs without exposing node features.

Core claim

At the paper's center is the claim that task, modality, and topology heterogeneity in federated multimodal graph learning are not separate nuisances but a single design problem, and that one mechanism can address all three at once. The mechanism is topology-aware cross-modal routing: each client computes per-modality prototypes by importance-weighted averaging with PageRank weights, so each prototype condenses both local semantics and graph structure; the server then compares these structure-informed prototypes across all clients and modalities, and for every client-modality pair retrieves the most similar other-client prototype as a positive and the least similar as a negative. These routed

What carries the argument

Topology-aware cross-modal prototype routing. Each client computes PageRank importance weights over its local graph, forms one prototype per modality (plus a fused-modality prototype) by weighted averaging of encoded node features, and uploads the prototypes under differential privacy. The server, for each client and each modality pair, selects the most cosine-similar prototype from another client as the positive contrastive reference and the least similar as the negative. These routed references feed a client-level contrastive loss that is added to node-level and neighbor-level contrastive losses, so the shared encoder is pulled toward cross-client, cross-modal alignment guided by local gra

Load-bearing premise

The load-bearing premise is that cosine similarity between prototypes from different clients is semantically meaningful in the shared embedding space despite the very modality and topology differences the method targets; if those prototypes are not comparable, the routed positives and negatives are arbitrary and the cross-client contrastive loss pulls local prototypes toward mismatched references.

What would settle it

Construct a federation with known ground-truth cross-client semantic correspondence (e.g., clients holding different views of the same items) and check whether Eq. (7) retrieves the true matching prototype as the positive more often than chance. If routing accuracy is near random on clients with high modality or topology heterogeneity, the client-level loss cannot be aligning what it claims to align.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A single task-agnostic encoder pre-trained across clients can be fine-tuned per client for node classification, link prediction, cross-modal retrieval, and graph-to-text or graph-to-image generation, so clients with different objectives can still collaborate.
  • Cross-client routing of prototypes transfers complementary modality knowledge; the ablation shows removing it causes the largest drop, e.g., 2.28% ACC on Movies and 4.70% R@1 on Toys.
  • Under differential privacy, the method retains over 99.4% of non-private performance at epsilon >= 8 and over 97.4% at epsilon = 1.
  • The method is robust to high feature, topology, and label sparsity, retaining 93-97% of accuracy at 50% sparsity.
  • Because pre-training is label-free, large parts of the pipeline work without downstream supervision.

Where Pith is reading between the lines

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

  • One prototype per modality likely underrepresents clients with several distinct semantic communities; extending routing to cluster-aware multi-prototype summaries is a natural next step and may improve expressiveness where a single centroid blurs community structure.
  • The routing step's usefulness hinges on cosine similarity being comparable across clients; a calibration stage or similarity normalization could be tested to see whether routed positives are genuinely closer than randomly chosen prototypes under high modality shift.
  • The two-stage paradigm suggests a progressive or continual pre-training variant that keeps updating the shared encoder during fine-tuning, potentially closing the gap between task-agnostic and task-specific objectives.
  • The same route-and-contrast pattern could transfer to other federated settings without graphs, such as private multimodal data spread across institutions with different sensing quality.

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 FedTCR, a two-stage federated multimodal graph learning method. In the pre-training stage, each client linearly projects per-modality features into a shared latent space, applies a low-pass graph filter, and constructs PageRank-weighted modality prototypes. The server aggregates encoder parameters and performs cross-modal prototype routing: for each client and modality pair it selects the most similar other-client prototype as a positive and the most dissimilar as a negative, using cosine similarity on DP-perturbed prototypes. These routed pairs drive a client-level contrastive loss added to node- and neighbor-level contrastive losses. After pre-training, each client fine-tunes an isolated task-specific head. Experiments cover node classification, link prediction, modality retrieval, and modality generation on 8 datasets across 7 domains, with ablations, robustness, hyperparameter sensitivity, and DP budget analyses.

Significance. If the central mechanism holds, FedTCR is a useful contribution to an emerging problem: it is one of the first systematic treatments of multimodal federated graph learning, it separates task-agnostic pre-training from task-specific fine-tuning, and it introduces a lightweight server-side routing scheme whose communication overhead is only a few prototype vectors per round. The paper is commendable for including an explicit pseudocode, a clean proof that Eq. (2) unifies SGC/APPNP-style filtering (Theorem 1), detailed dataset statistics, and a candid limitations section. The main empirical and methodological weaknesses are that the signature routing rule in Eq. (7) is not directly validated against alternative selection rules, the heterogeneous-task experiment is confounded by data scale, and hyperparameters appear to be selected on a test set that is also used for evaluation.

major comments (4)
  1. [Sec. 4.1, Eq. (7) and Table 2] The central novelty is the server-side argmax/argmin routing over cosine similarities of DP-perturbed, one-round-stale prototypes. The only ablation ('w/o CR') removes the entire client-level loss L_c, so it cannot distinguish the routing selection from any other prototype-based cross-client contrastive regularizer. If random routing or using all other prototypes as a uniform reference performed equally well, the 'topology-aware cross-modal routing' claim would be unsupported. Please add ablations that isolate the selection rule (e.g., random positive/negative, or averaging over all other prototypes) and ideally a diagnostic showing that the selected pairs have higher semantic correspondence than random pairs.
  2. [Sec. 5.3, Fig. 2] The task-heterogeneous experiment compares Isolated, FedTCR (homo., 4 clients) and FedTCR (hete., 20 clients). These arms differ not only in collaboration scope but also in the number of clients and the volume of pre-training data; the reported +2.44% average gain could be a data-scale effect rather than a task-heterogeneity effect. The statement that 'the arms differ only in the collaboration scope' is therefore inaccurate. A fairer control would hold the number of participants fixed or match the data budget. In addition, no external method is compared in this setting, so the abstract's 'state-of-the-art' claim is supported only for the homogeneous-task setting.
  3. [Sec. 5.6, Table 3] The hyperparameter sweep is reported on Movies, and Movies is also an evaluation dataset in Table 1. If the default hyperparameters were selected using Movies test-set metrics, the Movies row and the cross-domain 'consistent' claim are partially circular. Please use an internal validation split for model selection, or otherwise show that the default configuration was not chosen on the test metrics.
  4. [Sec. 4.1 and Algorithm 1] The protocol states that at each round the server samples a subset S of clients, but Eq. (7) and Algorithm 1 route among all K clients. With partial participation, the server would need cached prototypes from non-participating clients, and the DP/communication analysis in Appendix C would change. Experiments use full participation, but the algorithm description should either restrict routing to the sampled set or specify how stale prototypes are handled.
minor comments (4)
  1. [Eqs. (2) and (3)] The symbol α is used both as the smoothing strength in the graph filter (Eq. (2), α>0) and as the PageRank damping factor (Eq. (3), α∈(0,1)). These are different quantities; please rename one of them.
  2. [Appendix B.2] Eq. (4) says the PageRank vector is computed 'by power iteration' and 'after convergence', but the experimental configuration only lists 20 power iterations. Please specify the stopping tolerance or clarify that 20 iterations are used as a fixed budget.
  3. [Sec. 5.3] The claim 'enabling cross-task collective intelligence' is stronger than the evidence: the comparison in Fig. 2 shows a benefit over same-task federation, but does not attribute the benefit to task heterogeneity rather than to the larger number of clients. Consider softening the wording or adding a control.
  4. [Table 1] The caption states that 'N/A' indicates the algorithm is not applicable, but the reason differs across methods (e.g., FedGTA is label-dependent and cannot do link prediction, while unimodal FGL methods cannot do modality-centric tasks). A footnote explaining the different reasons would improve readability.

Circularity Check

0 steps flagged

No significant circularity: FedTCR's proposed routing and contrastive losses are internal training designs evaluated on held-out downstream metrics; self-citations are background, not load-bearing.

full rationale

The core derivation chain is self-contained. The multimodal encoder (Eqs. 1-2), PageRank prototypes (Eqs. 3-5), DP perturbation (Eq. 6), cross-modal routing (Eq. 7), and the three contrastive losses (Eqs. 8-10) are all specified in the paper and optimized on local/client objectives; no term in these equations is fitted to the downstream ACC/AUC/R@1/R-L/C-S numbers reported in Tables 1-2. The routing positive/negative pairs are training references selected by cosine similarity, and the client-level loss pulls local prototypes toward them; this is a learning mechanism, not a prediction that reduces to an input by construction. The claimed 'first systematic algorithm' positioning cites MM-OpenFGL (Li et al. 2026) and other prior work containing overlapping co-authors, but these citations are background/baseline context and are not used to prove the method's central claim or to exclude alternatives. Theorem 1 is a standard algebraic equivalence (SGC/GCN/APPNP as graph filters) proved in the appendix, not a borrowed uniqueness theorem. The main unaddressed issue—whether the cosine-based routing selects semantically meaningful cross-client pairs—is an empirical-validity concern, not circularity: the ablation shows removing the routing component hurts, and the claim would need a random-routing control, but the absence of that control does not make the derivation circular.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The method rests on standard graph signal processing and PageRank mathematics, plus domain assumptions about pre-trained encoders and the realism of Louvain partitions. The free parameters are hand-set defaults, several of which were selected using the Movies test set via the Table 3 sweep.

free parameters (7)
  • alpha_smoothing = 2.0
    Controls the geometric weighting in the graph filter Eq. (2); default chosen without a dedicated sensitivity sweep (Table 3 sweeps other hyperparameters).
  • pagerank_damping_alpha = 0.85
    Damping factor for PageRank in Eq. (3)-(4), standard default.
  • temperature_tau = 0.5
    Contrastive temperature in psi(u,v); tuned via the Movies sweep in Table 3.
  • loss_weights_lambda = lambda_m=0.3, lambda_g=0.2, lambda_c=0.5
    Balancing coefficients in Eq. (11); defaults are the best point found on the Movies sensitivity grid.
  • propagation_depth_L = 3
    Number of filter hops in Eq. (2); swept in Table 3, best at default 3.
  • dp_clipping_bound_C = 1.0
    L2 clipping bound in Eq. (6); set with delta=1e-5, epsilon varied in Sec. 5.7.
  • latent_dimension_d = 128
    Shared embedding dimension; set by hand, not swept.
axioms (5)
  • standard math Neumann series convergence of the geometric matrix filter (Eq. 12, Appendix F)
    Used to prove Theorem 1 and justify Eq. (2) as a low-pass filter.
  • standard math PageRank stationary distribution exists and power iteration converges for the client graphs
    Eq. (3)-(4) compute node importance weights.
  • domain assumption Frozen pre-trained encoders (Sentence-BERT, ViT, DINOv2, Qwen2-VL) provide features with sufficient cross-modal semantic alignment
    All experiments depend on this; stated in Appendix B.2 without validation.
  • domain assumption Louvain partitioning of each full MAG yields realistic client subgraphs
    Used to create the federated setting in Sec. 5.1.
  • standard math Gaussian mechanism with L2 clipping provides (epsilon,delta)-DP for the uploaded prototypes
    Eq. (6) and Sec. 5.7; standard DP result.

pith-pipeline@v1.3.0-alltime-deepseek · 25746 in / 13375 out tokens · 140944 ms · 2026-08-05T00:30:42.763499+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Towards Effective Federated Multimodal Graph Learning via Navigating Multifaceted Heterogeneity." pith.science (2026). https://pith.science/paper/72TFOUH4

@misc{pith2026260800623,
  author       = {Pith},
  title        = {Pith review of: Towards Effective Federated Multimodal Graph Learning via Navigating Multifaceted Heterogeneity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/72TFOUH4}},
  note         = {Machine review of arXiv:2608.00623}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Multimodal-attributed graphs (MAGs), where nodes carry heterogeneous semantic content across multiple modalities while edges encode relational dependencies, have been widely adopted across diverse domains. Federated multimodal graph learning (FMGL) extends federated graph learning (FGL) to MAGs, enabling collaborative optimization across decentralized MAGs without exposing raw data. However, naively applying existing FGL methods to FMGL is insufficient, as they fail to navigate the multifaceted heterogeneity inherent in decentralized MAGs, including task heterogeneity across diverse client objectives, modality heterogeneity from discrepant modality quality and semantic domains, and topology heterogeneity arising from divergent topological patterns with low cross-modality correlation. To address these challenges, we propose Federated multimodal graph learning with Topology-aware Cross-modal Routing (FedTCR), the first systematic algorithm designed for FMGL. To handle task heterogeneity, FedTCR employs a two-stage paradigm that comprises federated task-agnostic pre-training followed by isolated task-oriented fine-tuning. To jointly address modality and topology heterogeneity, FedTCR introduces a topology-aware cross-modal routing mechanism. Concretely, each client distills modality-specific knowledge into compact prototypes via topology-aware importance-weighted aggregation informed by graph structure; the server then evaluates cross-client cross-modal relationships among these structure-informed prototypes and routes informative ones as contrastive references, driving a tri-level cross-modal contrastive learning scheme that jointly aligns cross-client modalities while preserving discrimination. Experiments across 7 domains demonstrate that FedTCR outperforms state-of-the-art baselines on both graph-centric and modality-centric tasks.

Figures

Figures reproduced from arXiv: 2608.00623 by Di Wu, Guocong Quan, Miao Hu, Wang Luo, Wei-Jin Huang, Xunkai Li, Yinlin Zhu, Yi Zhang.

Figure 1
Figure 1. Figure 1: Overview of FedTCR, which addresses multifaceted heterogeneity in FMGL through federated task-agnostic pre [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance under the task-heterogeneous setting (4 clients for each dataset, i.e., one MAG per task). FedTCR [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Retained ACC on Movies under three sparsity [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] 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

127 extracted references · 35 canonical work pages · 1 internal anchor

  1. [1]

    Journal of Statistical Mechanics: Theory and Experiment , volume=

    Fast unfolding of communities in large networks , author=. Journal of Statistical Mechanics: Theory and Experiment , volume=. 2008 , publisher=

    2008

  2. [2]

    Nature Communications , volume=

    Biomedical knowledge graph learning for drug repurposing by extending guilt-by-association to multiple layers , author=. Nature Communications , volume=. 2023 , publisher=

    2023

  3. [3]

    IEEE Journal of Translational Engineering in Health and Medicine , year=

    A Graph Convolutional Network Based on Univariate Neurodegeneration Biomarker for Alzheimer’s Disease Diagnosis , author=. IEEE Journal of Translational Engineering in Health and Medicine , year=

  4. [4]

    Bioinformatics , volume=

    Similarity measures-based graph co-contrastive learning for drug--disease association prediction , author=. Bioinformatics , volume=. 2023 , publisher=

    2023

  5. [5]

    bioRxiv , pages=

    Ensemble-GNN: federated ensemble learning with graph neural networks for disease module discovery and classification , author=. bioRxiv , pages=. 2023 , publisher=

    2023

  6. [6]

    Expert Systems , pages=

    Federated learning-based private medical knowledge graph for epidemic surveillance in internet of things , author=. Expert Systems , pages=. 2023 , publisher=

    2023

  7. [7]

    Predictive Intelligence in Medicine: 5th International Workshop, PRIME 2022, Held in Conjunction with MICCAI 2022, Singapore, September 22, 2022, Proceedings , pages=

    Federated Time-Dependent GNN Learning from Brain Connectivity Data with Missing Timepoints , author=. Predictive Intelligence in Medicine: 5th International Workshop, PRIME 2022, Held in Conjunction with MICCAI 2022, Singapore, September 22, 2022, Proceedings , pages=. 2022 , organization=

    2022

  8. [8]

    Bioinformatics , volume=

    GNN-SubNet: disease subnetwork detection with explainable graph neural networks , author=. Bioinformatics , volume=. 2022 , publisher=

    2022

  9. [9]

    IEEE Journal of Biomedical and Health Informatics , year=

    Hyper-Graph Attention Based Federated Learning Method For Mental Health Detection , author=. IEEE Journal of Biomedical and Health Informatics , year=

  10. [10]

    Proceedings of the ACM Web Conference, WWW , pages=

    Graph-less Collaborative Filtering , author=. Proceedings of the ACM Web Conference, WWW , pages=

  11. [11]

    Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, WSDM , pages=

    DGRec: Graph Neural Network for Recommendation with Diversified Embedding Generation , author=. Proceedings of the Sixteenth ACM International Conference on Web Search and Data Mining, WSDM , pages=

  12. [12]

    International Conference on Learning Representations, ICLR , year=

    LightGCL: Simple Yet Effective Graph Contrastive Learning for Recommendation , author=. International Conference on Learning Representations, ICLR , year=

  13. [13]

    IEEE Transactions on Industrial Informatics , year=

    FGC: GCN based federated learning approach for trust industrial service recommendation , author=. IEEE Transactions on Industrial Informatics , year=

  14. [14]

    IEEE Transactions on Industrial Informatics , volume=

    Federated Learning-Based Cross-Enterprise Recommendation With Graph Neural Networks , author=. IEEE Transactions on Industrial Informatics , volume=. 2022 , publisher=

    2022

  15. [15]

    Vertical Federated Graph Neural Network for Recommender System

    Vertical Federated Graph Neural Network for Recommender System , author=. arXiv preprint arXiv:2303.05786 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  16. [16]

    Intelligent Systems with Applications , pages=

    Predicting Systemic Risk in Financial Systems Using Deep Graph Learning , author=. Intelligent Systems with Applications , pages=. 2023 , publisher=

    2023

  17. [17]

    Pacific-Asia Conference on Knowledge Discovery and Data Mining , pages=

    Anti-Money Laundering in Cryptocurrency via Multi-Relational Graph Neural Network , author=. Pacific-Asia Conference on Knowledge Discovery and Data Mining , pages=. 2023 , organization=

    2023

  18. [18]

    2023 IEEE 6th Information Technology, Networking, Electronic and Automation Control Conference , volume=

    Default Risk Assessment of Internet Financial Enterprises Based on Graph Neural Network , author=. 2023 IEEE 6th Information Technology, Networking, Electronic and Automation Control Conference , volume=. 2023 , organization=

    2023

  19. [19]

    2022 IEEE International Conference on Cloud Computing Technology and Science (CloudCom) , pages=

    2SFGL: A Simple And Robust Protocol For Graph-Based Fraud Detection , author=. 2022 IEEE International Conference on Cloud Computing Technology and Science (CloudCom) , pages=. 2022 , organization=

    2022

  20. [20]

    ACM Computing Surveys , volume=

    AI in Finance: Challenges, Techniques, and Opportunities , author=. ACM Computing Surveys , volume=. 2022 , publisher=

    2022

  21. [21]

    IEEE Transactions on Dependable and Secure Computing , year=

    Your labels are selling you out: Relation leaks in vertical federated learning , author=. IEEE Transactions on Dependable and Secure Computing , year=

  22. [22]

    Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, KDD , pages=

    Optuna: A next-generation hyperparameter optimization framework , author=. Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, KDD , pages=

  23. [23]

    2012 , publisher=

    Matrix analysis , author=. 2012 , publisher=

    2012

  24. [24]

    International conference on machine learning , pages=

    Scaffold: Stochastic controlled averaging for federated learning , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  25. [25]

    Applied Sciences , volume=

    Fedopt: Towards communication efficiency and privacy preservation in federated learning , author=. Applied Sciences , volume=. 2020 , publisher=

    2020

  26. [26]

    arXiv preprint arXiv:1912.00818 , year=

    Federated learning with personalization layers , author=. arXiv preprint arXiv:1912.00818 , year=

    Pith/arXiv arXiv 1912

  27. [27]

    Knowledge-Based Systems , volume=

    A survey on federated learning , author=. Knowledge-Based Systems , volume=. 2021 , publisher=

    2021

  28. [28]

    International journal of machine learning and cybernetics , volume=

    A survey on federated learning: challenges and applications , author=. International journal of machine learning and cybernetics , volume=. 2023 , publisher=

    2023

  29. [29]

    Artificial intelligence and statistics , pages=

    Communication-efficient learning of deep networks from decentralized data , author=. Artificial intelligence and statistics , pages=. 2017 , organization=

    2017

  30. [30]

    International conference on theory and applications of models of computation , pages=

    Differential privacy: A survey of results , author=. International conference on theory and applications of models of computation , pages=. 2008 , organization=

    2008

  31. [31]

    Proceedings of the 2016 ACM SIGSAC conference on computer and communications security , pages=

    Deep learning with differential privacy , author=. Proceedings of the 2016 ACM SIGSAC conference on computer and communications security , pages=

    2016

  32. [32]

    Proceedings of Machine learning and systems , volume=

    Federated optimization in heterogeneous networks , author=. Proceedings of Machine learning and systems , volume=

  33. [33]

    arXiv preprint arXiv:2404.14061 , year=

    Fedtad: Topology-aware data-free knowledge distillation for subgraph federated learning , author=. arXiv preprint arXiv:2404.14061 , year=

    Pith/arXiv arXiv

  34. [34]

    International conference on machine learning , pages=

    Personalized subgraph federated learning , author=. International conference on machine learning , pages=. 2023 , organization=

    2023

  35. [35]

    Advances in neural information processing systems , volume=

    Subgraph federated learning with missing neighbor generation , author=. Advances in neural information processing systems , volume=

  36. [36]

    Proceedings of the VLDB Endowment , year=

    FedGTA: Topology-Aware Averaging for Federated Graph Learning , author=. Proceedings of the VLDB Endowment , year=

  37. [37]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Federated graph-level clustering network , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  38. [38]

    Forty-second International Conference on Machine Learning , year=

    Federated node-level clustering network with cross-subgraph link mending , author=. Forty-second International Conference on Machine Learning , year=

  39. [39]

    Neural Networks , pages=

    Federated graph-level clustering network with adaptive knowledge compensation , author=. Neural Networks , pages=. 2025 , publisher=

    2025

  40. [40]

    2024 , eprint=

    Federated Graph Semantic and Structural Learning , author=. 2024 , eprint=

    2024

  41. [41]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Federated graph learning under domain shift with generalizable prototypes , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  42. [42]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Federated learning on non-iid graphs via structural knowledge sharing , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  43. [43]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Modeling inter-intra heterogeneity for graph federated learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  44. [44]

    2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    FedSPA : Generalizable Federated Graph Learning under Homophily Heterogeneity , author=. 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    2025

  45. [45]

    arXiv preprint arXiv:2507.02409 , year=

    S2FGL: Spatial Spectral Federated Graph Learning , author=. arXiv preprint arXiv:2507.02409 , year=

    Pith/arXiv arXiv

  46. [46]

    Advances in neural information processing systems , volume=

    Federated graph classification over non-iid graphs , author=. Advances in neural information processing systems , volume=

  47. [47]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Model-contrastive federated learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  48. [48]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Feddc: Federated learning with non-iid data via local drift decoupling and correction , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  49. [49]

    arXiv preprint arXiv:2506.01973 , year=

    Multimodal financial foundation models (MFFMs): Progress, prospects, and challenges , author=. arXiv preprint arXiv:2506.01973 , year=

    Pith/arXiv arXiv

  50. [50]

    BMC bioinformatics , volume=

    A multimodal graph neural network framework for cancer molecular subtype classification , author=. BMC bioinformatics , volume=. 2024 , publisher=

    2024

  51. [51]

    Sentence-bert: Sentence embeddings using siamese bert-networks , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

  52. [52]

    Journal of machine learning research , volume=

    Exploring the limits of transfer learning with a unified text-to-text transformer , author=. Journal of machine learning research , volume=

  53. [53]

    arXiv preprint arXiv:2006.03677 , year=

    Visual transformers: Token-based image representation and processing for computer vision , author=. arXiv preprint arXiv:2006.03677 , year=

    Pith/arXiv arXiv 2006

  54. [54]

    arXiv preprint arXiv:2304.07193 , year=

    Dinov2: Learning robust visual features without supervision , author=. arXiv preprint arXiv:2304.07193 , year=

    Pith/arXiv arXiv

  55. [55]

    2020 , eprint=

    Iterative Deep Graph Learning for Graph Neural Networks: Better and Robust Node Embeddings , author=. 2020 , eprint=

    work page 2020

  56. [56]

    Proceedings of the ACM on Web Conference 2025 , pages=

    Unigraph2: Learning a unified embedding space to bind multimodal graphs , author=. Proceedings of the ACM on Web Conference 2025 , pages=

    2025

  57. [57]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Lgmrec: Local and global graph learning for multimodal recommendation , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

    work page

  58. [58]

    arXiv preprint arXiv:2310.07478 , year=

    Multimodal graph learning for generative tasks , author=. arXiv preprint arXiv:2310.07478 , year=

    Pith/arXiv arXiv

  59. [59]

    Advances in Neural Information Processing Systems , volume=

    Instructg2i: Synthesizing images from multimodal attributed graphs , author=. Advances in Neural Information Processing Systems , volume=

    work page

  60. [60]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Graphgpt-o: Synergistic multimodal comprehension and generation on graphs , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    work page

  61. [61]

    Information Processing & Management , volume=

    Mgat: Multimodal graph attention network for recommendation , author=. Information Processing & Management , volume=. 2020 , publisher=

    2020

  62. [62]

    Proceedings of the 27th ACM international conference on multimedia , pages=

    MMGCN: Multi-modal graph convolution network for personalized recommendation of micro-video , author=. Proceedings of the 27th ACM international conference on multimedia , pages=

  63. [63]

    Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    MGNN: A multimodal graph neural network for predicting the survival of cancer patients , author=. Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

    work page

  64. [64]

    Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , pages=

    Multimodal graph transformer for multimodal question answering , author=. Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics , pages=

    work page

  65. [65]

    ACM Transactions on Multimedia Computing, Communications and Applications , volume=

    Multimodal graph for unaligned multimodal sequence analysis via graph convolution and graph pooling , author=. ACM Transactions on Multimedia Computing, Communications and Applications , volume=. 2023 , publisher=

    work page 2023

  66. [66]

    arXiv preprint arXiv:2511.20030 , year=

    Cross-contrastive clustering for multimodal attributed graphs with dual graph filtering , author=. arXiv preprint arXiv:2511.20030 , year=

    arXiv

  67. [67]

    arXiv preprint arXiv:2601.22416 , year=

    MM-OpenFGL: A Comprehensive Benchmark for Multimodal Federated Graph Learning , author=. arXiv preprint arXiv:2601.22416 , year=

    arXiv

  68. [68]

    International conference on multimedia modeling , pages=

    Sentiment analysis on multi-view social data , author=. International conference on multimedia modeling , pages=. 2016 , organization=

    work page 2016

  69. [69]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    A suite of generative tasks for multi-level multimodal webpage understanding , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    work page 2023

  70. [70]

    Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval , pages=

    Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning , author=. Proceedings of the 44th international ACM SIGIR conference on research and development in information retrieval , pages=

  71. [71]

    ECCV Workshops (2) , series=

    How to Read Paintings: Semantic Art Understanding with Multi-modal Retrieval , author=. ECCV Workshops (2) , series=

    work page

  72. [72]

    Proceedings of the IEEE International Conference on Computer Vision, ICCV , year=

    Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models , author=. Proceedings of the IEEE International Conference on Computer Vision, ICCV , year=

  73. [73]

    arXiv preprint arXiv:2403.03952 , year=

    Bridging language and items for retrieval and recommendation , author=. arXiv preprint arXiv:2403.03952 , year=

    Pith/arXiv arXiv

  74. [74]

    Proceedings of the 12th ACM conference on recommender systems , pages=

    Item recommendation on monotonic behavior chains , author=. Proceedings of the 12th ACM conference on recommender systems , pages=

    work page

  75. [75]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    Fine-grained spoiler detection from large-scale review corpora , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    work page

  76. [76]

    Advances in Neural Information Processing Systems, NeurIPS, Datasets and Benchmarks Track, NeurIPS DB Track , year=

    RedCaps: Web-curated image-text data created by the people, for the people , author=. Advances in Neural Information Processing Systems, NeurIPS, Datasets and Benchmarks Track, NeurIPS DB Track , year=

  77. [77]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    Ninerec: A benchmark dataset suite for evaluating transferable recommendation , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  78. [78]

    Justifying recommendations using distantly-labeled reviews and fine-grained aspects , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

  79. [79]

    IEEE transactions on neural networks and learning systems , volume=

    A comprehensive survey on graph neural networks , author=. IEEE transactions on neural networks and learning systems , volume=. 2020 , publisher=

    2020

  80. [80]

    AI Open , volume=

    Graph neural networks: A review of methods and applications , author=. AI Open , volume=. 2020 , publisher=

    2020

Showing first 80 references.