arxiv: 2605.11919 · v1 · submitted 2026-05-12 · 💻 cs.LG

Recognition: 2 theorem links

· Lean Theorem

STAGE: Tackling Semantic Drift in Multimodal Federated Graph Learning

Guoren Wang, Rong-Hua Li, Xunkai Li, Xun Wu, Yihan Sun, Zekai Chen

Pith reviewed 2026-05-13 07:38 UTC · model grok-4.3

classification 💻 cs.LG

keywords multimodal federated graph learningsemantic driftsemantic calibrationgraph message passingfederated learningmultimodal attributesinconsistency regulation

0 comments

The pith

STAGE translates heterogeneous multimodal features into a shared semantic space and regulates their propagation over local graphs to reduce inconsistency in federated learning.

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

The paper targets the core difficulty in multimodal federated graph learning: clients holding different modalities produce inconsistent representations for the same concepts, so direct parameter averaging creates false agreements and graph message passing spreads those errors. STAGE solves this by inserting an explicit translation step that maps the mismatched features into comparable vectors, then adds a regulation mechanism that controls how the aligned vectors travel across each client's local graph. The result is improved cross-client calibration without forcing a naive shared space upfront. Experiments across eight multimodal-attributed graphs and five tasks show higher accuracy than prior methods while cutting per-round communication volume.

Core claim

STAGE is a protocol-first framework that first translates heterogeneous multimodal features into comparable representations and then regulates how these representations propagate over local graph structures, thereby improving cross-client semantic calibration and reducing the risk of inconsistency amplification during graph learning.

What carries the argument

The translation step that maps heterogeneous multimodal features into a comparable semantic space, combined with subsequent regulation of graph message passing on local structures.

If this is right

Cross-client semantic calibration improves because representations start in an aligned space rather than being forced together after the fact.
Inconsistency amplification is limited because regulation controls how aligned vectors spread through graph neighborhoods.
State-of-the-art accuracy holds across both graph-centric tasks such as node classification and modality-centric tasks such as cross-modal retrieval.
Per-round communication payload drops because the protocol avoids exchanging raw high-dimensional multimodal parameters.
The same protocol works on eight different multimodal-attributed graphs without task-specific redesign.

Where Pith is reading between the lines

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

The translation-plus-regulation pattern could be tested in non-graph federated settings where clients hold mismatched data modalities.
If residual inconsistencies remain after translation, adaptive regulation weights might further reduce error spread on sparse graphs.
Privacy budgets could be tightened because less raw multimodal data needs to be exchanged once semantic alignment is in place.
The approach suggests semantic alignment is a reusable primitive for any distributed multimodal system that must avoid drift.

Load-bearing premise

A translation step can reliably map heterogeneous multimodal features into a comparable semantic space without introducing false agreements or losing task-relevant information, and regulating graph propagation will prevent amplification of any residual inconsistency.

What would settle it

Running STAGE on a new multimodal-attributed graph where accuracy stays below standard federated averaging baselines or per-round communication payload does not decrease would falsify the central performance claim.

Figures

Figures reproduced from arXiv: 2605.11919 by Guoren Wang, Rong-Hua Li, Xunkai Li, Xun Wu, Yihan Sun, Zekai Chen.

**Figure 1.** Figure 1: Observation 1: Semantic drift. Even for the same semantic concept, clients from different modality domains can produce clearly separated representations, leading to large centroid drift in a shared diagnostic space. comparability: clients may refer to the same concept, yet their learned representations are not directly aligned. Consider the category running shoes. A text-dominant client may emphasize words… view at source ↗

**Figure 2.** Figure 2: Observation 2: Pseudo-alignment. A naively shared anchor bank does not guarantee shared semantics: different clients may map different meanings to the same anchor. site fails, improving server-side aggregation alone is inherently insufficient to recover true semantic consistency across clients. A seemingly natural remedy is to introduce a shared semantic space across clients. Yet in MM-FGL, naive anchor s… view at source ↗

**Figure 5.** Figure 5: Validation of pseudo-alignment correction. Delta heatmap of anchor-class assignment distributions. The heatmap in [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗

**Figure 4.** Figure 4: Validation of pseudo-alignment correction. Dominant-class purity gain across top activated anchors. b) Stage II: Pseudo-Alignment as the Failure of Naive Shared Anchors.: A natural response to semantic drift is to impose a shared semantic space, e.g., through a frozen anchor bank. However, this alone does not ensure true agreement. If incompatible local features are projected into a shared anchor space wit… view at source ↗

**Figure 7.** Figure 7: Systematic miscalibration of local attention. (Left) A centralized oracle clearly separates homophilous and heterophilous edges. (Right) In the federated setting, limited structural exposure causes severe overlap between the two distributions. tially, with the overall mean drift growing by 29.8%. This confirms that once semantic inconsistency enters message passing, it becomes a graph-level amplification e… view at source ↗

**Figure 8.** Figure 8: Overview of the STAGE framework. The protocol-first architecture decouples federated graph learning into client execution and server evolution. Local [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗

**Figure 9.** Figure 9: Robustness against feature drift. Performance decay of different methods across four datasets as feature drift intensity (α) increases from 0.0 to 1.0. STAGE exhibits superior resilience, whereas parameter-centric and uncalibrated multimodal baselines suffer from severe degradation. 2 5 10 20 30 Noise Ratio (%) 75.5 76.0 76.5 77.0 77.5 78.0 ACC (%) Node Classification on Toys FedAvg FedLap FedMVP STAGE (Ou… view at source ↗

**Figure 10.** Figure 10: Robustness against modality noise on QB, Bili Music, Toys, Flickr30k, and SemArt datasets. We randomly replace the input text and image [PITH_FULL_IMAGE:figures/full_fig_p008_10.png] view at source ↗

**Figure 11.** Figure 11: Hyperparameter sensitivity. (Left) Relative performance drop under different semantic calibration weights λ. (Right) Heatmap over semantic calibration and entropy regularization hyperparameters (λ, β). Hyperparameter Sensitivity.We perform a sensitivity analysis on key hyperparameters in STAGE. For the InfoNCE semantic calibration (Lgap), which resolves feature drift, we vary its strength coefficient λ (… view at source ↗

**Figure 12.** Figure 12: Scalability across Client Partitions. Performance impact of scaling the number of decentralized clients (K ∈ {3, 5, 7}). TABLE IV EFFICIENCY AND RESOURCE COMPARISON Method Comm. Payload Total Ops Space Cost Time (Scalars) (FLOPs) (MB) (s) FedAvg 1.00 × 106 1.7 × 108 44.9 0.0891 FedMAC 1.28 × 105 3.1 × 108 41.2 0.1599 FedSPA 1.00 × 106 3.6 × 109 45.8 1.8257 STAGE (Ours) 8.19 × 103 7.3 × 108 6.2 0.4225 coor… view at source ↗

**Figure 13.** Figure 13: Efficiency Analysis on Bili Music. (Left) STAGE achieves faster convergence and a higher AUC ceiling compared to baselines. (Right) Total computational time demonstrates that STAGE adds only marginal overhead to FedAvg while being significantly faster than complex FGL methods. Accelerated Training Convergence [PITH_FULL_IMAGE:figures/full_fig_p009_13.png] view at source ↗

read the original abstract

Federated graph learning (FGL) enables collaborative training on graph data across multiple clients. As graph data increasingly contain multimodal node attributes such as text and images, multimodal federated graph learning (MM-FGL) has become an important yet substantially harder setting. The key challenge is that clients from different modality domains may not share a common semantic space: even for the same concept, their local encoders can produce inconsistent representations before collaboration begins. This makes direct parameter coordination unreliable and further causes two downstream problems: forcing heterogeneous client representations into a naively shared semantic space may create false semantic agreement, and graph message passing may amplify residual inconsistency across neighborhoods. To address this issue, we propose \textbf{STAGE}, a protocol-first framework for MM-FGL. Instead of relying on direct parameter averaging, STAGE builds a shared semantic space that first translates heterogeneous multimodal features into comparable representations and then regulates how these representations propagate over local graph structures. In this way, STAGE not only improves cross-client semantic calibration, but also reduces the risk of inconsistency amplification during graph learning. Extensive experiments on 8 multimodal-attributed graphs across 5 graph-centric and modality-centric tasks show that STAGE consistently achieves state-of-the-art performance while reducing per-round communication payload.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

STAGE translates multimodal features into a shared space then regulates graph propagation to cut semantic drift in federated settings, claiming SOTA results and lower communication, but without direct checks on whether the translation preserves signal or avoids false matches.

read the letter

The main point is that STAGE builds a shared semantic space by translating heterogeneous multimodal node features first, then regulates how those representations propagate on each client's local graph. This is positioned as better than direct parameter averaging when encoders start from inconsistent representations for the same concepts. The two problems it targets—false agreements from forced sharing and amplification through message passing—are described clearly and seem like real obstacles in multimodal federated graph learning. Experiments on eight multimodal-attributed graphs across five tasks report consistent gains plus reduced per-round communication, which would matter for practical privacy-preserving setups if the numbers hold with solid controls. The protocol emphasis on explicit translation plus regulation looks distinct from the simpler alignment tricks mentioned in the abstract. The soft spot is the lack of direct evidence on the translation step itself. No alignment diagnostics appear in the reported results, such as retrieval accuracy between original and translated embeddings or checks for lost task-relevant information, so the performance edge could trace to regularization effects or dataset quirks rather than reliable calibration. The abstract also stays light on exact baseline details and quantitative deltas, which makes it harder to judge robustness without the full tables and ablations. This paper is aimed at people working on federated graph methods with mixed node attributes like text and images. A reader focused on deployment issues in heterogeneous settings would find the mechanism worth examining. It has enough of a concrete idea and experimental breadth to deserve serious referee time, though it will probably need added diagnostics on the translation quality to land cleanly.

Referee Report

2 major / 1 minor

Summary. The paper proposes STAGE, a protocol-first framework for multimodal federated graph learning (MM-FGL) that tackles semantic drift by first translating heterogeneous multimodal node features (e.g., text and images) from different clients into a shared semantic space and then regulating how these representations propagate over local graph structures, rather than relying on direct parameter averaging. It claims this improves cross-client semantic calibration, reduces inconsistency amplification during graph message passing, and yields state-of-the-art performance on 8 multimodal-attributed graphs across 5 graph-centric and modality-centric tasks while also lowering per-round communication payload.

Significance. If the central claims hold after proper validation, the work would be significant for the growing area of multimodal federated graph learning, as it directly targets the mismatch in semantic spaces across modality-specific encoders and the risk of error amplification in graph propagation, potentially enabling more reliable collaborative training with reduced communication overhead.

major comments (2)

[Abstract] Abstract: The central performance claims (SOTA results and communication savings across 8 graphs and 5 tasks) are asserted without any quantitative metrics, baseline details, ablation studies, or statistical significance tests. This absence prevents verification of whether the reported gains are attributable to the proposed translation-plus-regulation mechanism or to other factors such as regularization effects or dataset biases.
[Method] Method section (translation and regulation components): The manuscript provides no direct diagnostics for the translation step's fidelity, such as cross-modal retrieval accuracy, mutual information between original and translated embeddings on matched concepts, or measures of false semantic agreement. Without these, it is impossible to confirm that the shared semantic space avoids information loss or spurious alignments, which is load-bearing for the claim that subsequent graph regulation prevents inconsistency amplification.

minor comments (1)

[Abstract] The abstract and introduction could more explicitly define the five tasks and the precise communication payload metric (e.g., bytes per round or number of parameters exchanged) to allow readers to assess the savings claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments. We address each point below and have revised the manuscript to strengthen the presentation of results and validation of the translation component.

read point-by-point responses

Referee: [Abstract] Abstract: The central performance claims (SOTA results and communication savings across 8 graphs and 5 tasks) are asserted without any quantitative metrics, baseline details, ablation studies, or statistical significance tests. This absence prevents verification of whether the reported gains are attributable to the proposed translation-plus-regulation mechanism or to other factors such as regularization effects or dataset biases.

Authors: The abstract serves as a concise overview, while the full manuscript details all quantitative metrics, baselines, ablations, and statistical tests in the Experiments section. To address the concern, we will revise the abstract to include key numerical results (e.g., average accuracy gains and communication reductions) along with references to the supporting tables and significance tests. This change will make the claims immediately verifiable without altering the underlying evidence that the gains stem from the proposed mechanisms rather than regularization or biases. revision: yes
Referee: [Method] Method section (translation and regulation components): The manuscript provides no direct diagnostics for the translation step's fidelity, such as cross-modal retrieval accuracy, mutual information between original and translated embeddings on matched concepts, or measures of false semantic agreement. Without these, it is impossible to confirm that the shared semantic space avoids information loss or spurious alignments, which is load-bearing for the claim that subsequent graph regulation prevents inconsistency amplification.

Authors: We agree that explicit diagnostics would strengthen validation of the translation step. In the revised manuscript we will add cross-modal retrieval accuracy and mutual information analyses between original and translated embeddings, reported in the Experiments section. These will demonstrate fidelity of the shared space and support that regulation addresses residual inconsistencies rather than relying on potentially spurious alignments. revision: yes

Circularity Check

0 steps flagged

No circularity detected in derivation chain

full rationale

The paper introduces STAGE as a protocol-first framework that translates heterogeneous multimodal features into a shared semantic space and then regulates their propagation over local graphs to mitigate semantic drift in MM-FGL. No equations, fitted parameters, or self-referential definitions are described that would reduce the claimed improvements to redefinitions of the inputs by construction. The central mechanisms are presented as novel additions rather than tautological mappings, and performance is asserted via experiments on external benchmarks rather than internal consistency alone. Any self-citations are not load-bearing for the core claims.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the domain assumption that heterogeneous encoders produce inconsistent representations for the same concept and that a translation mechanism can create a usable shared space without external supervision.

axioms (1)

domain assumption Clients from different modality domains produce inconsistent representations for the same concept before collaboration.
Explicitly stated as the key challenge in the abstract.

invented entities (1)

Shared semantic space constructed via translation no independent evidence
purpose: To make heterogeneous multimodal features comparable across clients
Core component of STAGE introduced to solve semantic drift.

pith-pipeline@v0.9.0 · 5530 in / 1167 out tokens · 33649 ms · 2026-05-13T07:38:37.135351+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
Variational semantic translation... qv = arg min ... + τs DKL(q∥u) ... max-entropy L(k)ent ... Global Anchor Prototypes ... meta-controller πθπ(Dk)
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
STAGE consistently achieves state-of-the-art performance while reducing per-round communication payload

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages

[1]

Fedgraphnn: A federated learning system and benchmark for graph neural networks,

C. He, K. Balasubramanian, E. Ceyani, C. Yang, H. Xie, L. Sun, L. He, L. Yang, P. S. Yu, Y . Ronget al., “Fedgraphnn: A federated learning system and benchmark for graph neural networks,”arXiv preprint arXiv:2104.07145, 2021

work page arXiv 2021
[2]

A comprehensive data-centric overview of federated graph learning,

Z. Wu, X. Li, Y . Zhu, Z. Chen, G. Yan, Y . Yan, H. Zhang, Y . Ai, X. Jin, R.-H. Liet al., “A comprehensive data-centric overview of federated graph learning,”arXiv preprint arXiv:2507.16541, 2025

work page arXiv 2025
[3]

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

X. Li, Y . Ai, Y . Zhu, H. Lu, Y . Zhang, G. Fu, B. Fan, Q. Dai, R.- H. Li, and G. Wang, “Mm-openfgl: A comprehensive benchmark for multimodal federated graph learning,”arXiv preprint arXiv:2601.22416, 2026

work page arXiv 2026
[4]

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,” inArtificial intelligence and statistics. Pmlr, 2017, pp. 1273– 1282

work page 2017
[5]

Federated optimization in heterogeneous networks,

T. Li, A. K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, and V . Smith, “Federated optimization in heterogeneous networks,”Proceedings of Machine learning and systems, vol. 2, pp. 429–450, 2020

work page 2020
[6]

Scaffold: Stochastic controlled averaging for federated learn- ing,

S. P. Karimireddy, S. Kale, M. Mohri, S. Reddi, S. Stich, and A. T. Suresh, “Scaffold: Stochastic controlled averaging for federated learn- ing,” inInternational conference on machine learning. PMLR, 2020, pp. 5132–5143

work page 2020
[7]

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, vol. 36, no. 8, 2022, pp. 8432–8440

work page 2022
[8]

Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning,

J. Zhang, Y . Liu, Y . Hua, and J. Cao, “Fedtgp: Trainable global prototypes with adaptive-margin-enhanced contrastive learning for data and model heterogeneity in federated learning,” inProceedings of the AAAI conference on artificial intelligence, vol. 38, no. 15, 2024, pp. 16 768–16 776

work page 2024
[9]

Multimodal feder- ated learning via contrastive representation ensemble,

Q. Yu, Y . Liu, Y . Wang, K. Xu, and J. Liu, “Multimodal feder- ated learning via contrastive representation ensemble,”arXiv preprint arXiv:2302.08888, 2023

work page arXiv 2023
[10]

Federated graph classification over non-iid graphs,

H. Xie, J. Ma, L. Xiong, and C. Yang, “Federated graph classification over non-iid graphs,”Advances in neural information processing sys- tems, vol. 34, pp. 18 839–18 852, 2021. JOURNAL OF LATEX CLASS FILES, VOL. XX, NO. X, MAY 2026 10

work page 2021
[11]

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. PMLR, 2023, pp. 1396–1415

work page 2023
[12]

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,”arXiv preprint arXiv:2401.11755, 2024

work page arXiv 2024
[13]

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,” arXiv preprint arXiv:2404.14061, 2024

work page arXiv 2024
[14]

A meta-computing framework for collaborative federated graph learning in industrial iots,

X. Zheng, X. Hu, T. Wang, Q. Huang, and L. Zhang, “A meta-computing framework for collaborative federated graph learning in industrial iots,” IEEE Internet of Things Journal, 2025

work page 2025
[15]

Rethinking client- oriented federated graph learning,

Z. Chen, X. Li, Y . Zhu, R.-H. Li, and G. Wang, “Rethinking client- oriented federated graph learning,” inProceedings of the 34th ACM International Conference on Information and Knowledge Management, 2025, pp. 393–402

work page 2025
[16]

Ensemble distillation for robust model fusion in federated learning,

T. Lin, L. Kong, S. U. Stich, and M. Jaggi, “Ensemble distillation for robust model fusion in federated learning,”Advances in neural information processing systems, vol. 33, pp. 2351–2363, 2020

work page 2020
[17]

Federated mutual learning,

T. Shen, J. Zhang, X. Jia, F. Zhang, G. Huang, P. Zhou, K. Kuang, F. Wu, and C. Wu, “Federated mutual learning,”arXiv preprint arXiv:2006.16765, 2020

work page arXiv 2006
[18]

Fedmultimodal: A bench- mark for multimodal federated learning,

T. Feng, D. Bose, T. Zhang, R. Hebbar, A. Ramakrishna, R. Gupta, M. Zhang, S. Avestimehr, and S. Narayanan, “Fedmultimodal: A bench- mark for multimodal federated learning,” inProceedings of the 29th ACM SIGKDD conference on knowledge discovery and data mining, 2023, pp. 4035–4045

work page 2023
[19]

Semi- supervised graph classification: A hierarchical graph perspective,

J. Li, Y . Rong, H. Cheng, H. Meng, W. Huang, and J. Huang, “Semi- supervised graph classification: A hierarchical graph perspective,” inThe World Wide Web Conference, 2019, pp. 972–982

work page 2019
[20]

Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks,

J. Lu, D. Batra, D. Parikh, and S. Lee, “Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks,”Advances in neural information processing systems, vol. 32, 2019

work page 2019
[21]

Lion: Empowering multimodal large language model with dual-level visual knowledge,

G. Chen, L. Shen, R. Shao, X. Deng, and L. Nie, “Lion: Empowering multimodal large language model with dual-level visual knowledge,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26 540–26 550

work page 2024
[22]

Higpt: Heterogeneous graph language model,

J. Tang, Y . Yang, W. Wei, L. Shi, L. Xia, D. Yin, and C. Huang, “Higpt: Heterogeneous graph language model,” inProceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining, 2024, pp. 2842–2853

work page 2024
[23]

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,” inProceedings of the ACM on Web Conference 2025, 2025, pp. 1759–1770

work page 2025
[24]

Justifying recommendations using distantly-labeled reviews and fine-grained aspects,

J. Ni, J. Li, and J. McAuley, “Justifying recommendations using distantly-labeled reviews and fine-grained aspects,” inProceedings of the 2019 conference on empirical methods in natural language pro- cessing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), 2019, pp. 188–197

work page 2019
[25]

Ninerec: A benchmark dataset suite for evaluating transferable recommendation,

J. Zhang, Y . Cheng, Y . Ni, Y . Pan, Z. Yuan, J. Fu, Y . Li, J. Wang, and F. Yuan, “Ninerec: A benchmark dataset suite for evaluating transferable recommendation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

work page 2024
[26]

Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models,

B. A. Plummer, L. Wang, C. M. Cervantes, J. C. Caicedo, J. Hocken- maier, and S. Lazebnik, “Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 2641– 2649

work page 2015
[27]

How to read paintings: semantic art un- derstanding with multi-modal retrieval,

N. Garcia and G. V ogiatzis, “How to read paintings: semantic art un- derstanding with multi-modal retrieval,” inProceedings of the European Conference on Computer Vision (ECCV) Workshops, 2018, pp. 0–0

work page 2018
[28]

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

work page 2008
[29]

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

work page 1998
[30]

Inductive representation learning on large graphs,

W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,”Advances in neural information processing systems, vol. 30, 2017

work page 2017
[31]

Graph attention networks,

P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y . Bengio et al., “Graph attention networks,” inInternational conference on learning representations, vol. 6, no. 2. Ithaca, 2018

work page 2018
[32]

Leveraging foundation models for multi-modal federated learning with incomplete modality,

L. Che, J. Wang, X. Liu, and F. Ma, “Leveraging foundation models for multi-modal federated learning with incomplete modality,” inJoint European Conference on Machine Learning and Knowledge Discovery in Databases. Springer, 2024, pp. 401–417

work page 2024
[33]

Fedmac: Tackling partial-modality missing in feder- ated learning with cross-modal aggregation and contrastive regulariza- tion,

M. D. Nguyen, T. T. Nguyen, H. H. Pham, T. N. Hoang, P. Le Nguyen, and T. T. Huynh, “Fedmac: Tackling partial-modality missing in feder- ated learning with cross-modal aggregation and contrastive regulariza- tion,” in2024 22nd International Symposium on Network Computing and Applications (NCA). IEEE, 2024, pp. 278–285

work page 2024
[34]

Subgraph federated learning via spectral methods,

J. Aliakbari, J. ¨Ostman, A. Panahiet al., “Subgraph federated learning via spectral methods,”arXiv preprint arXiv:2510.25657, 2025

work page arXiv 2025
[35]

S2fgl: Spatial spectral federated graph learning,

Z. Tan, S. Huang, G. Wan, W. Huang, H. Li, and M. Ye, “S2fgl: Spatial spectral federated graph learning,”arXiv preprint arXiv:2507.02409, 2025

work page arXiv 2025
[36]

Fedspa: Generalizable federated graph learning under homophily het- erogeneity,

Z. Tan, G. Wan, W. Huang, H. Li, G. Zhang, C. Yang, and M. Ye, “Fedspa: Generalizable federated graph learning under homophily het- erogeneity,” inProceedings of the Computer Vision and Pattern Recog- nition Conference, 2025, pp. 15 464–15 475

work page 2025
[37]

Modeling inter-intra heterogeneity for graph federated learning,

W. Yu, S. Chen, Y . Tong, T. Gu, and C. Gong, “Modeling inter-intra heterogeneity for graph federated learning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 21, 2025, pp. 22 236– 22 244. Zekai Chenis currently pursuing his Master’s degree in Computer Science at Beijing Institute of Technology under the supervision of Prof...

work page 2025
[38]

Her research interests include computer science and technology

She is currently a second-year undergraduate student majoring in Computer Science and Technol- ogy at Minzu University of China, Beijing, China. Her research interests include computer science and technology. Rong-Hua Lireceived the PhD degree from the Chinese University of Hong Kong, in 2013. He is currently a professor with the Beijing Institute of Tech...

work page 2013