Pith. sign in

REVIEW 4 major objections 5 minor 37 references

FedOGL claims catastrophic forgetting in federated open-world multimodal graph learning is a semantic-structural memory failure, and its three-part framework cuts forgetting by 42.67% while preserving or improving downstream accuracy.

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 03:41 UTC pith:MYO26OO7

load-bearing objection A solid empirical framework for a genuinely new federated open-world graph learning setting, with real gains over baselines, but reproducibility gaps and an unproven subspace assumption keep it from being a clean accept. the 4 major comments →

arxiv 2607.27665 v1 pith:MYO26OO7 submitted 2026-07-30 cs.LG

FedOGL: Combating Catastrophic Forgetting in Federated Open-World Multimodal Graph Learning

Zekai Chen , Haodong Lu , Shihao Li , Weiwei Ji , Xunkai Li , Xun Wu , Yinlin Zhu , Rong-Hua Li This is my paper
classification cs.LG
keywords federated graph learningcatastrophic forgettingopen-world learningmultimodal graphscontinual learningout-of-distribution rejectiongraph neural networksprototype consolidation
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.

FedOGL tackles catastrophic forgetting in federated open-world multimodal graph learning, where multiple clients hold private graph data and must keep learning new classes over time while remembering old ones and flagging unknown samples. The paper argues that forgetting in this setting is not just a classifier-level failure but a semantic-structural memory failure driven by three mechanisms: new classes overwrite cross-modal semantics, new-task gradients erode the graph-propagation directions old classes depend on, and client-wise partial data fragments category memory across the federation. FedOGL counters each mechanism: replay plus task-start distillation anchors historical decision behavior, a globally aggregated structural basis shields structural gradients from overwriting old propagation directions, and a server-side prototype memory matches and merges fragmented category evidence. Across six datasets and two tasks, FedOGL reports cutting forgetting (FM) by 42.67% and unknown-acceptance (FPR95) by 28.31% relative to the best baseline, while keeping or improving final accuracy. If correct, it provides a unified way to keep distributed models both plastic to new classes and stable to old ones.

Core claim

FedOGL's core discovery is that catastrophic forgetting in federated open-world multimodal graphs is a semantic-structural memory failure with three identifiable causes, and that each cause has a corresponding remedy that can be composed into one framework. Modality-semantic overwriting is countered by anchoring the current model to a frozen task-start teacher and replaying small historical batches. Topology-induced structural erosion is countered by maintaining a global structural basis—built by aggregating per-client residual projectors—and projecting structural gradients orthogonal to that basis, so updates that would rewrite historical propagation directions are suppressed. Federated mem

What carries the argument

The load-bearing object is the global structure basis U_t: clients extract residual structural directions from centered embeddings (subtracting what the previous basis already explains), send projector matrices U_k,t U_k,t^T to the server, which sums them to form S_t and takes the top eigenvectors as the new global basis. In the next stage, structural gradients are multiplied by (I - U_t U_t^T), the orthogonal projection, so gradient components that would move parameters along protected historical directions are zeroed. The second key object is the server-side prototype memory bank: local semantic-structural descriptors per class are uploaded, matched, and merged with support-weighted averag

Load-bearing premise

The whole result rests on the structural subspace U_t capturing precisely the graph-propagation directions whose preservation keeps old classes alive; if that basis is incomplete—because local embeddings are modality-deficient or the current stage's data does not expose historical directions—the shielding either misses what needs protecting or blocks what needs learning.

What would settle it

Construct a task stream where historical classes are supported by propagation directions that do not appear in the current stage's top residual eigenvectors (e.g., the old class is only reachable through a modality that the current clients lack), then train FedOGL with replay and distillation intact; if FM still degrades as much as a no-shielding baseline, the structure basis is not carrying the memory it claims to protect.

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

If this is right

  • Federated graph models can keep learning new classes indefinitely without a globally synchronized class catalog, as long as each client retains a bounded replay buffer and the server maintains the structure basis and prototype bank.
  • The structural-gradient projection operates on gradient geometry rather than model architecture, so it should transfer to any graph encoder with a structural representation pathway.
  • Unknown rejection improves as a side effect of retention: because old-class boundaries are preserved, the maximum-softmax confidence rule misclassifies fewer OOD samples as known.
  • Server-side prototype consolidation provides a privacy-preserving way to share category knowledge across asynchronous clients, since only class indices and compact descriptors cross the network.
  • FedOGL retains early-task knowledge throughout the incremental trajectory, not just at the final stage, as shown by the task-wise retention analysis on Grocery.

Where Pith is reading between the lines

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

  • The projector-summing aggregation in Eq. (8) is rotation-insensitive: summing projectors rather than vectors means the global basis does not depend on arbitrary rotations of local coordinate systems, which could make the shielding robust to client drift in ways that direct averaging would not—but the paper does not test this explicitly.
  • Because the structure basis is built from embeddings at the current stage, it can only protect directions that are still visible in the representation; classes whose structural signature has already been erased before the basis is computed would be missed. A testable extension is to estimate U_t from a mixture of current and replayed embeddings.
  • The prototype-graph refinement (Eqs. 13–14) is essentially a tiny graph continual-learning problem on the server; hooking it into class-balanced sampling or more advanced OOD detection could further reduce FPR95 without extra client cost.
  • The claim that forgetting is a semantic-structural memory failure suggests a diagnostic: monitor D_sem, D_str, and D_fed during training and check whether FedOGL's gains track all three; if only D_str drops while D_sem stays high, the three-mechanism decomposition may be incomplete.

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 / 5 minor

Summary. The paper formulates federated open-world multimodal graph learning under class-incremental evolution, arguing that catastrophic forgetting in this setting arises from three mechanisms: modality-semantic overwriting, topology-induced structural erosion, and federated memory fragmentation. It proposes FedOGL, which combines task-anchored semantic retention (replay + distillation), structure-basis memory shielding (global structural subspace projection), and prototype-guided memory consolidation (server-side prototype matching/merging and prototype-graph refinement). Experiments on six datasets across node classification and cross-modal retrieval report that FedOGL outperforms FL/FGL/open-world baselines, with headline claims of a 42.67% FM reduction and a 28.31% FPR95 reduction.

Significance. If the results hold, this is a useful new benchmark and a practically strong framework for a realistic decentralized incremental setting. The paper introduces a clear three-mechanism diagnostic decomposition, reports five-seed means and standard deviations, includes ablations and sensitivity analysis, and proposes a distinctive structural-subspace protection idea. The strength of the empirical claims is, however, reduced by two issues: the headline improvement percentages are not computed against the best-performing baseline in Table 1, and the central structural-basis assumption is not directly validated. These concerns currently prevent acceptance.

major comments (4)
  1. [Abstract; Table 1; Section 5.1] The reported '42.67% FM reduction' is not the improvement over the best-performing baseline. Averaging FM over the six datasets gives FedOGL 5.24% and FedIIH 9.14%, i.e., (9.14-5.24)/9.14 = 42.67%. But FedIIH is not the strongest baseline: on RedditS, POWER achieves FM 3.49% versus FedOGL's 5.64%, and on KU, CLIPN has a lower FM (12.89%) than FedIIH (18.76%). Similarly, '28.31% FPR95 reduction' equals the reduction from POWER's average FPR95 (45.89% to 32.90%), but on RedditS, KU, and QB the best baseline is TopoOOD or CLIPN, not POWER. Please recompute the headline statistics against the true best baseline per dataset or explicitly define the comparison basis.
  2. [Sec. 4.2, Eqs. (7)-(9)] The structure-basis shielding assumes that U_t captures the graph-propagation directions whose preservation prevents old-class forgetting. This is not established. Eq. (7) builds residual bases from current-stage embeddings after removing U_{t-1}; historical directions not already in U_{t-1} and not re-expressible in current/replayed embeddings are irrecoverable, and the fixed rank r in Eq. (8) can drop old eigenvectors. Eq. (9) suppresses updates aligned with the leading variance directions of the current embedding covariance, but there is no argument that high-variance directions are exactly the memory-critical directions. The ablation in Table 2 shows that removing the projection hurts, but a generic gradient regularizer could produce the same pattern. Please add a direct test: measure the overlap between the protected basis U_t and the historical task bases U_i used in the D_str diag
  3. [Sec. 3, Eqs. (2)-(4)] The three-mechanism decomposition is supported only by descriptive drift metrics that are not tied to downstream task performance. No evidence is given that reducing D_sem, D_str, or D_fed is causally responsible for the FM improvements, rather than being a correlate of generic regularization. Since the paper's conceptual contribution is the three-mechanism interpretation, this connection needs support, e.g., a correlation analysis between each drift metric and FM across methods, or intervention experiments that isolate each mechanism.
  4. [Experimental Setup; Reproducibility] Essential implementation details are deferred to 'Section 2 of the supporting material' and 'Section 4 of the supporting material', but no supplementary material is included with the manuscript. Critical choices — the OOD confidence threshold delta_t in Eq. (6), the prototype matching threshold eta and memory slots S in Eq. (12), the basis rank r, loss weights, and replay memory size M — are not specified in the main text. Without these details, the five-seed results cannot be audited or reproduced. Please include all such details in the submission or state explicitly that the supplement is part of the reviewed version.
minor comments (5)
  1. [Abstract; Introduction] The abstract contains an unrelated passage ('As cammermay splice an authentic interview clip...') and a duplicated sentence ('In this setting, clients must learn emerging classes...'). The Introduction also contains a blank reference 'As illustrated in Sec. .' These text corruptions must be fixed before publication.
  2. [Eq. (7)] The class-balanced sample weights W_{k,t} are not defined. Specify their dimension and computation.
  3. [Eq. (6)] The OOD confidence threshold delta_t is a key free parameter, but its selection protocol is not described. The sensitivity analysis in Fig. 7 covers only basis rank and distillation weight, not delta_t, eta, or memory size.
  4. [Figures 5 and 7] Several figure labels appear garbled or use nonstandard character encodings, making the axis and legend text difficult to read. Please regenerate the figures with standard fonts.
  5. [References] Several references are dated 2026 and marked 'arXiv preprint'. If these are not yet publicly available or accepted, please clarify their status or remove them.

Circularity Check

0 steps flagged

No significant circularity: the headline forgetting reduction is an empirical comparison against external baselines on standard metrics, not a prediction forced by the paper's own definitions or fitted parameters.

full rationale

FedOGL's central claim is an empirical result: FM, FPR95, and AM are reported against three categories of external baselines (FL, FGL, open-world) on standard continual-learning protocols. The three forgetting mechanisms (modality-semantic overwriting, topology-induced structural erosion, federated memory fragmentation) are introduced through paper-defined drift diagnostics (Eqs. 2–4), and the three modules are designed to reduce those drift measures; this is an internally consistent design narrative rather than a circular prediction, because the headline 42.67% FM reduction is not computed from those diagnostics but from task-wise performance drops relative to baselines. The structure-basis shielding (Eqs. 7–9) rests on an unproven assumption that top eigenvectors of current structural embeddings span historical graph-propagation directions; that is a robustness/correctness risk, not a circularity, since the claimed FM advantage is not equivalent to the projection operation by construction. Self-citations to STAGE, PRISM, and MM-OpenFGL appear in related work and in dataset/baseline selection, but they are not used as load-bearing evidence for FedOGL's effectiveness, and no uniqueness theorem or ansatz is imported from the authors' prior work. No fitted parameter is renamed as a prediction. The skeptical concern about U_t omitting historical directions is a plausible limitation, but the paper does not define its success metric in terms of U_t, so the derivation chain is not circular.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim depends heavily on hand-chosen thresholds, ranks, loss weights, and replay sizes, none of which are fully reported in the submitted text. The mathematical projection steps are standard linear algebra, but the semantic assumptions about what the structural basis contains and what the prototypes represent are domain assumptions that are not independently verified.

free parameters (5)
  • OOD confidence threshold delta_t = not reported
    Eq. (6) uses delta_t to decide OOD versus known class; FPR95 and all unknown-rejection results depend directly on it. No value or tuning schedule is given in the submitted text.
  • Structural basis ranks r and r_l (TopEig_r) = not reported
    Eqs. (7)-(8) truncate the local residual basis and global structural basis to top eigenvectors; this rank controls how much structural gradient is shielded and is tuned by hand. Figure 7 studies sensitivity but the reported values are not stated.
  • Prototype matching threshold eta and memory bank slots S = not reported
    Eq. (12) merges an incoming prototype only when cosine similarity exceeds eta and maintains S slots per class; both determine how cross-client memory is consolidated and are not reported.
  • Loss weights lambda_rep, lambda_dis, lambda_lab, lambda_tr, temperature tau, schedule gamma_t = not reported
    Eqs. (5), (13), and (14) balance new-class plasticity, replay, distillation, and server-side prototype-graph refinement. Their values affect the headline AM/FM/FPR95 numbers and are not given in the text.
  • Replay memory size M = not reported
    Eq. (5) replays q = min(M, t-1) previous stages. The bounded replay buffer is a primary forgetting-mitigation lever, and its size is not stated for the reported experiments.
axioms (5)
  • domain assumption Clients can retain a bounded local replay buffer of raw historical labeled nodes.
    Eq. (5) requires replay over historical stages. The privacy claim is that raw data is not shared, but local retention of old raw samples is assumed; if data-retention is disallowed, the replay component and a large part of the method fail.
  • domain assumption Each client observes only a subset of new classes, and unknown classes never appear during training.
    Eq. (1) defines the evolving known/unknown class space with C_U appearing only in evaluation; this is the formal definition of the problem setting, not an externally verified property of real multimodal graph streams.
  • domain assumption Louvain-partitioned five-client subgraphs are a representative surrogate for real non-IID federated multimodal graphs.
    The experimental protocol partitions six public datasets with Louvain and creates class-incremental streams; the realism of this proxy for actual federated deployments is assumed.
  • domain assumption Maximum softmax probability is a sufficiently calibrated OOD score and delta_t can be set to hold 95% TPR.
    Eq. (6) relies on the MSP baseline of Hendrycks and Gimpel; the reported FPR95 depends on threshold calibration that is not documented.
  • standard math Projector summation and residual projection preserve the subspace geometry needed for structural memory shielding.
    Eqs. (7)-(9) use standard linear-algebra projection onto top-eigenvector subspaces; this is mathematically valid but whether the resulting subspace contains the right semantic-structural memory is assumed.
invented entities (1)
  • Prototype graph G_P = (P_t, E_P) with same-class and KNN edges no independent evidence
    purpose: Server-side consolidation substrate: the global model is refined on this graph of class prototypes plus trajectory-aware soft targets.
    This is an algorithm-introduced object rather than an external physical entity. Its utility is supported only by the paper's ablations; no independent evidence validates that prototype-graph refinement preserves semantics better than alternative consolidation mechanisms.

pith-pipeline@v1.3.0-daily-deepseek · 13984 in / 10923 out tokens · 117027 ms · 2026-08-01T03:41:50.969734+00:00 · methodology

0 comments
read the original abstract

Federated graph learning enables collaborative training over decentralized graph data without sharing raw graph information. As such risks evolve, clients must learn emerging classes from private multimodal graph streams, retain historical categories, and reject samples outside the known class space. In this setting, clients must learn emerging classes from private multimodal graph streams while preserving historical categories and rejecting samples outside the current known class space. The core challenge is catastrophic forgetting, which in federated multimodal graphs is not merely a classifier-level failure: old knowledge can be erased through modality-semantic overwriting, topology-induced structural erosion, and federated memory fragmentation. To address this challenge, we propose \textbf{FedOGL}, a semantic-structural memory preservation framework. On the client side, FedOGL preserves historical decision behavior through replay and task-start distillation, while protecting graph-propagation memory via projection onto a globally shared structure basis. On the server side, FedOGL maintains and transfers compact category prototypes to facilitate cross-client knowledge sharing without exposing raw graph data. Extensive experiments demonstrate that, compared with the best-performing baselines, FedOGL reduces performance degradation caused by catastrophic forgetting by \textbf{42.67\%}, while maintaining or improving performance on downstream tasks.

Figures

Figures reproduced from arXiv: 2607.27665 by Haodong Lu, Rong-Hua Li, Shihao Li, Weiwei Ji, Xunkai Li, Xun Wu, Yinlin Zhu, Zekai Chen.

Figure 1
Figure 1. Figure 1: Federated open-world multimodal graph learning [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Client-to-global prototype discrepancies for Fe [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of FedOGL. At each incremental stage, clients learn from private multimodal graph streams and exchange only [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Federation-scale analysis. The two panels show [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: AM and FM sensitivity to the structure-basis rank [PITH_FULL_IMAGE:figures/full_fig_p007_7.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

37 extracted references · 8 linked inside Pith

  1. [1]

    Artificial intelligence and statistics , pages=

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

    2017

  2. [2]

    Proceedings of Machine learning and systems , volume=

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

  3. [3]

    ICLR 2021 Workshop on Distributed and Private Machine Learning (DPML) , year=

    Fedgraphnn: A federated learning benchmark system for graph neural networks , author=. ICLR 2021 Workshop on Distributed and Private Machine Learning (DPML) , year=

    2021

  4. [4]

    arXiv preprint arXiv:2408.16288 , year=

    Openfgl: A comprehensive benchmark for federated graph learning , author=. arXiv preprint arXiv:2408.16288 , year=

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2401.11755 , year=

    Fedgta: Topology-aware averaging for federated graph learning , author=. arXiv preprint arXiv:2401.11755 , year=

    Pith/arXiv arXiv

  6. [6]

    Advances in Neural Information Processing Systems , volume=

    FedSSP: Federated Graph Learning with Spectral Knowledge and Personalized Preference , author=. Advances in Neural Information Processing Systems , volume=

  7. [7]

    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=. 2025 , doi=

    2025

  8. [8]

    arXiv preprint arXiv:2605.11919 , year=

    STAGE: Tackling Semantic Drift in Multimodal Federated Graph Learning , author=. arXiv preprint arXiv:2605.11919 , year=

    Pith/arXiv arXiv

  9. [9]

    arXiv preprint arXiv:2606.09301 , year=

    PRISM: Topology-Aware Cross-Modal Imputation for Modality-Deficient Federated Graph Learning , author=. arXiv preprint arXiv:2606.09301 , year=

    Pith/arXiv arXiv

  10. [10]

    Chen, Zekai and Lu, Haodong and Li, Xunkai and Sun, Henan and Li, Jia and Qin, Hongchao and Li, Rong-Hua and Wang, Guoren , booktitle =

  11. [11]

    Proceedings of the national academy of sciences , volume=

    Overcoming catastrophic forgetting in neural networks , author=. Proceedings of the national academy of sciences , volume=. 2017 , publisher=

    2017

  12. [12]

    IEEE transactions on pattern analysis and machine intelligence , volume=

    Learning without forgetting , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2018 , publisher=

    2018

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    CGLB: Benchmark Tasks for Continual Graph Learning , author=. Advances in Neural Information Processing Systems , volume=

  14. [14]

    Proceedings of the 40th International Conference on Machine Learning , series=

    Towards Robust Graph Incremental Learning on Evolving Graphs , author=. Proceedings of the 40th International Conference on Machine Learning , series=

  15. [15]

    Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , pages=

    icarl: Incremental classifier and representation learning , author=. Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , pages=

  16. [16]

    arXiv preprint arXiv:2108.00158 , year=

    Multiplex graph networks for multimodal brain network analysis , author=. arXiv preprint arXiv:2108.00158 , year=

    Pith/arXiv arXiv

  17. [17]

    Proceedings of the 38th International Conference on Machine Learning , series=

    Learning Transferable Visual Models From Natural Language Supervision , author=. Proceedings of the 38th International Conference on Machine Learning , series=

  18. [18]

    International Conference on Learning Representations , year=

    Adam: A Method for Stochastic Optimization , author=. International Conference on Learning Representations , year=

  19. [19]

    arXiv preprint arXiv:1807.03748 , year=

    Representation Learning with Contrastive Predictive Coding , author=. arXiv preprint arXiv:1807.03748 , year=

    Pith/arXiv arXiv

  20. [20]

    International Conference on Learning Representations , year=

    A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks , author=. International Conference on Learning Representations , year=

  21. [21]

    2023 , doi =

    Zhang, Jie and Chen, Chen and Zhuang, Weiming and Lyu, Lingjuan , booktitle =. 2023 , doi =

    2023

  22. [22]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    Federated continual graph learning , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 , pages=

  23. [23]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    FedMVP: Federated Multimodal Visual Prompt Tuning for Vision-Language Models , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  24. [24]

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

    Towards open set deep networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  25. [25]

    Advances in neural information processing systems , volume=

    Energy-based out-of-distribution detection , author=. Advances in neural information processing systems , volume=

  26. [26]

    Forty-first International Conference on Machine Learning , year=

    Graph out-of-distribution detection goes neighborhood shaping , author=. Forty-first International Conference on Machine Learning , year=

  27. [27]

    Advances in Neural Information Processing Systems , volume=

    Revisiting score propagation in graph out-of-distribution detection , author=. Advances in Neural Information Processing Systems , volume=

  28. [28]

    arXiv preprint arXiv:2505.13989 , year=

    When LLMs meet open-world graph learning: a new perspective for unlabeled data uncertainty , author=. arXiv preprint arXiv:2505.13989 , year=

    Pith/arXiv arXiv

  29. [29]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Clipn for zero-shot ood detection: Teaching clip to say no , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  30. [30]

    arXiv preprint arXiv:2408.11237 , year=

    Out-of-Distribution Detection with Attention Head Masking for Multimodal Document Classification , author=. arXiv preprint arXiv:2408.11237 , year=

    Pith/arXiv arXiv

  31. [31]

    Advances in neural information processing systems , volume=

    Delving into out-of-distribution detection with vision-language representations , author=. Advances in neural information processing systems , volume=

  32. [32]

    arXiv preprint arXiv:2601.22416 , year=

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

    arXiv

  33. [33]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Mosaic of modalities: A comprehensive benchmark for multimodal graph learning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  34. [34]

    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

  35. [35]

    2023 , publisher=

    Wang, Xiaodan and Wang, Chengyu and Li, Lei and Li, Zhixu and Chen, Ben and Jin, Linbo and Huang, Jun and Xiao, Yanghua and Gao, Ming , booktitle=. 2023 , publisher=

    2023

  36. [36]

    2019 , publisher=

    Wei, Yinwei and Wang, Xiang and Nie, Liqiang and He, Xiangnan and Hong, Richang and Chua, Tat-Seng , booktitle=. 2019 , publisher=

    2019

  37. [37]

    Journal of statistical mechanics: theory and experiment , volume=

    Fast unfolding of communities in large networks , author=. Journal of statistical mechanics: theory and experiment , volume=