acceptodds
Under review as a conference paper at ICLR 2027

Towards Effective Federated Multimodal Graph Learning via Navigating Multifaceted Heterogeneity

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 FedTCR (Federated multimodal graph learning with Topology-aware Cross-modal Routing), 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. Extensive experiments across 6 domains demonstrate that FedTCR achieves substantial improvements over state-of-the-art baselines on both graph-centric and modality-centric tasks.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal-attributed Graph Learning, Federated Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WTuBOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper identifies federated multimodal graph learning (FMGL), where clients privately hold multimodal-attributed graphs, and argues that existing federated graph learning methods fail to navigate three forms of heterogeneity, namely task, modality, and topology heterogeneity. It proposes FedTCR, which uses a two-stage paradigm of federated task-agnostic pre-training followed by isolated task-oriented fine-tuning to handle task heterogeneity. During pre-training, each client encodes modalities into a shared space with a low-pass graph filter (Eq. 1 to 2), builds per-modality prototypes through PageRank importance-weighted aggregation (Eq. 3 to 5), and uploads them with differentially private perturbation (Eq. 6). The server routes the most and least similar cross-client cross-modal prototypes as positive and negative references (Eq. 7), driving a tri-level contrastive scheme at node, neighbor, and client levels (Eq. 8 to 11). Experiments on 8 MAG datasets across 7 domains cover graph-centric and modality-centric tasks under homogeneous and heterogeneous task settings, with ablations, sparsity robustness, hyperparameter sensitivity, and a privacy-utility study.

Critical strengths
  1. The problem decomposition into task, modality, and topology heterogeneity is clear and well motivated. Table 1 shows the issue is practical rather than hypothetical, since federated methods with an MMGCN backbone fall below isolated training on several columns (FedAvg 49.07 versus MMGCN 53.47 on Movies ACC, and 68.96 versus 81.37 on ele-fashion R@1), while FedTCR recovers and exceeds isolated performance on all eight datasets.

  2. The design maps onto the stated challenges in a clean way, and the ablation in Table 2 supports each component. Every removal degrades performance, with w/o CR causing the largest drop (-4.70% R@1 on Toys) and w/o TS confirming that task-agnostic pre-training matters.

  3. The evaluation covers more ground than is typical for this area. Eight datasets in seven domains, both graph-centric and modality-centric tasks including G2Text and G2Image, homogeneous and heterogeneous task settings, sparsity robustness (Figure 3), and a DP study with usable performance at epsilon 1 (Table 4).

Critical weaknesses
  1. Some margins are small relative to the reported variance, which makes the per-dataset conclusions hard to judge. On Grocery, FedTCR scores 81.01 ACC against 80.93 for Isolated training (std 0.40), and on Movies the gain over MMGCN is 0.80 points (54.27 versus 53.47).

  2. The heterogeneous-task experiment in Section 5.3 is a self-comparison, since existing methods cannot join such a federation. Because each task is instantiated on a different MAG, the +2.44% average gain of FedTCR (hete.) over FedTCR (homo.) in Figure 2 mixes task heterogeneity with dataset diversity.

  3. The privacy treatment is stated only at the prototype level. DP is applied to the uploaded prototypes (Eq. 6) while encoder parameters are shared without protection, and the per-round budget and composition behind the epsilon values in Table 4 are not given.

Questions for authors

Refer to weaknesses.

1 comment

Thank you for these constructive comments. We address your concerns as follows:


W1. Small margins relative to variance. Thank you for your constructive feedback. We interpret these cells through paired per-seed differences rather than mean rankings alone. Using the same five data splits/seeds for each method, we report the mean paired difference, its standard deviation, and the two-sided 95% t-based confidence interval (four degrees of freedom):

ComparisonFedTCRComparatorΔ\DeltaPaired std95% CIInterpretation
Grocery vs Isolated81.0180.93+0.080.29[-0.28,+0.44]Comparable
Movies vs MMGCN54.2753.47+0.800.21[+0.54,+1.06]Positive

Grocery is comparable, whereas Movies shows a positive paired difference whose 95% CI lies above zero. The broader performance conclusion rests on the trend across eight datasets and the larger margins on other tasks. For Movies, the gain is 0.80 percentage points, corresponding to an approximately 1.50% relative improvement over MMGCN.


W2. Task heterogeneity versus dataset diversity. Thank you for your constructive feedback. The original Figure 2 evaluates collaboration across task/dataset pairs. To distinguish task composition from dataset diversity, we compared same- and mixed-task federations on the same Movies graph and 12 partitions, using auxiliary downstream heads during pre-training. Same-task federations are trained separately for NC, LP, and MR; the mixed-task federation combines these objectives. Collaboration scale and budget remain unchanged:

FederationNC ACCLP AUCMR R@1
Same-task53.68$\pm$0.3294.86$\pm$0.2751.74$\pm$0.43
Mixed-task54.02$\pm$0.2995.31$\pm$0.2452.83$\pm$0.37

Since only task composition differs, this comparison separates its effect from dataset diversity; all three metrics favor mixed-task pre-training. We will scope Figure 2 to collaboration across task/dataset pairs.


W3. Privacy scope and composition. We apologize for this ambiguity. The formal DP guarantee covers prototype uploads; raw data remain local, while encoder parameters are shared without DP and can leak information. We will revise the abstract/conclusion to make this scope explicit. We will specify the privacy unit, neighboring relation, sensitivity, per-release budget, δ\delta, clipping, and cumulative composition over 600 releases per client (3 × 200 rounds). Table 4 evaluates prototype-channel privacy–utility only; parameter sharing remains outside this guarantee.

Reviewer_Jd28OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedTCR to address task, modality, and topology heterogeneity in federated multimodal graph learning. It adopts a two-stage framework consisting of federated task-agnostic pre-training and local task-oriented fine-tuning. Specifically, a topology-aware cross-modal routing mechanism is introduced to construct structure-informed prototypes and select cross-client contrastive references, supported by a tri-level contrastive learning strategy.

Critical strengths
  • The paper addresses multiple forms of heterogeneity in federated multimodal graph learning
  • The proposed FedTCR employs a well-designed contrastive learning strategy
  • The method achieves strong experimental results across diverse domains
Critical weaknesses
  • FedTCR mainly integrates existing techniques, including task-agnostic pre-training, PageRank-based prototype construction, and contrastive learning. The methodological novelty beyond their combination is not sufficiently clear.
  • FedTCR uses shared PageRank weights across modalities, which primarily capture node importance rather than modality-specific topology differences. It is unclear how this design effectively addresses topology heterogeneity across modalities.
  • FedTCR selects positive and negative prototypes based on cosine similarities. However, high similarity does not necessarily indicate semantic consistency across heterogeneous clients, similar to the least similar prototypes.
  • Each client constructs only one prototype per modality. Is that sufficient to capture complex semantic distributions and community structures within multimodal graphs? In addition, although differential privacy is applied to shared prototypes, model parameters are also communicated during federated training. The privacy implications of parameter sharing and cumulative privacy loss require further investigation.
  • The contrastive learning design in Eqs. (8–10) lacks detailed justification and analysis. For instance, the distinction between easy and hard positive/negative samples or compare different sampling strategies to validate the effectiveness of their design.
  • The paper lacks comprehensive empirical analysis of computational overhead, communication costs, and scalability as the number of clients increases, which is important for federated learning scenarios.
Questions for authors

See above Weakens.

1 comment

Thank you for these constructive comments.


W1. Methodological novelty.

We respectfully argue that technical novelty lies in how a design answers its motivations. Task-agnostic pre-training is established in multi-task learning, and PageRank is a classical graph-learning technique; their use as components does not imply limited technical novelty. FedTCR assigns them distinct roles: task-agnostic pre-training and isolated fine-tuning support diverse client objectives; per-modality prototypes and personalized cross-client, cross-modal routing exchange knowledge beyond parameter averaging; structure-informed prototypes and node/neighbor objectives preserve node importance and local topology.

FedTCR is not a direct transfer from adjacent fields: unimodal FGL lacks modality-specific knowledge and objectives; centralized MGL has no private routing; graph-agnostic multimodal FL lacks topology-informed prototypes. Our technical contribution is coordinated routing and tri-level optimization addressing all three challenges, rather than PageRank or contrastive learning alone.


W2. Shared PageRank weights.

We apologize for this ambiguity. Since all modalities share AkA_k, PageRank is modality-agnostic within a client and client-specific across clients. It weights nodes for prototypes; routing and tri-level objectives address modality-dependent semantics. We compared uniform, shared-graph, and modality-specific semantic-graph PageRank:

VariantMovies ACCRedditS AUCToys R@1
Uniform53.69$\pm$0.3096.89$\pm$0.2152.45$\pm$0.42
Shared PageRank54.27$\pm$0.2597.84$\pm$0.1853.64$\pm$0.36
Modality-specific PageRank54.08$\pm$0.2997.91$\pm$0.2353.51$\pm$0.40

Shared PageRank outperforms Uniform throughout; modality-specific weighting adds no consistent gain.


W3. Semantic reliability of routing.

Thank you for your constructive feedback. We evaluate routing reliability through downstream performance and label-distribution agreement. Eq. (7) selects max-similarity positives and min-similarity (easy) negatives; Eq. (10) imposes their contrastive ordering.

Positive / negativeMovies ACCRedditS AUCToys R@1
Max / min (Full)54.27$\pm$0.2597.84$\pm$0.1853.64$\pm$0.36
Random / random53.41$\pm$0.3596.73$\pm$0.2752.28$\pm$0.45
Max / hard negative53.29$\pm$0.3796.51$\pm$0.2952.06$\pm$0.48

We also measure query-reference label-distribution cosine agreement:

ReferenceMoviesGrocery
Routed positive0.82$\pm$0.040.79$\pm$0.05
Random reference0.61$\pm$0.070.58$\pm$0.09
Easy negative0.43$\pm$0.110.40$\pm$0.13
Hard negative0.72$\pm$0.060.68$\pm$0.08

Routed positives have the strongest label agreement, while hard negatives agree more than easy negatives, indicating greater false-negative risk for hard-negative sampling.


W4a. Single-prototype representation.

Thank you for your constructive feedback. Prototypes guide routing; node/neighbor objectives retain instance and local-topology variation. Cluster-aware prototypes yield:

qqMovies ACCRedditS AUCSemArt C-S
1 (default)54.27$\pm$0.2597.84$\pm$0.1869.45$\pm$0.22
254.39$\pm$0.2797.92$\pm$0.2069.72$\pm$0.31
Communication qq1248
KiB1544155115661596

q=1q=1 is the efficient default; q=2q=2 offers small gains at minimal additional cost.


W4b. Privacy scope and composition.

We apologize for this ambiguity. Prototype uploads have a formal DP guarantee; raw data remain local, but encoder parameters are shared without DP and can leak information. We will clarify this scope and report the privacy unit, neighboring relation, sensitivity, clipping, δ\delta, per-release budget, and cumulative composition over 600 releases per client (3 × 200 rounds). Table 4 evaluates prototype-channel privacy–utility only; parameter sharing is outside this guarantee.


W5. Justification of Eqs. (8)–(10).

Thank you for your constructive feedback. Eq. (8) contrasts cross-modal views of each node with impostors. Eq. (9) uses random-walk neighbors as positives and samples nodes outside this positive set as negatives. Eq. (10) uses routed references for cross-client transfer. Table 2 supports LmL_m and LgL_g; the sampling comparison in W3 evaluates LcL_c.


W6. Overhead, communication, and scalability.

Thank you for your constructive feedback. PageRank is one-time preprocessing; fine-tuning has no communication. Each client uploads three prototypes and receives twelve routed vectors per round, adding 7.5 KiB (0.49%) to 1.50 MiB of encoder traffic. Same-task results with fixed data/budget:

KKTime (s/round)Routing (ms)Toys R@1
54.960.0353.64$\pm$0.36
205.370.4153.58$\pm$0.34
806.716.3753.46$\pm$0.44

Across K=5,20,80K=5,20,80, routing time remains at millisecond scale with approximately quadratic growth, while Toys R@1 changes by only 0.18 points.

Reviewer_QtjJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper provided a solution for Federated multimodal graph learning, where clients with privacy text, image and node graph, to collaboratively train a multimodal encoder for downstream tasks, by exchanging model parameters and prototypes. This paper reveal three challenges such as task heterogeneity, modality heterogeneity and topology heterogeneity. It proposed FedTCR method, where federated pretraining was carried out between server and clients, following by downstream task finetuning locally on each clients. During the pretraining, each client provided its own local prototype, which was generated by aggregated text/image vectors with the weights from PageRank, while on the server side, encoders were aggregated by FedAvg, and the cross-model prototype routing was carried out, where the local prototypes from different clients were paired by cosine similarity. Based on the pairing, the positive and negative prototypes were selected for the following three levels contrastive loss: node-level, neighbor-level and client-level which related to the three challenges mentioned before. Last, proposed FedTCR were evaluated on 8 different datasets including Node Classification, link prediction, retrieval as well as generative tasks. The main results reported proposed method outperforming ove baselines and ablation study also demonstrated the proposed mechanism.

Critical strengths
  1. The core mechanism of FedTCR was demonstrated by the solid evaluation results: (1) FedTCR gain uplift over baselines on 8 benchmarks and different kinds of tasks. (2) After remove the prototype routing as well as the client-level contrastive loss, the AUC dropped from 97.84 to 95.67 on RedditS and R@1 dropped from 53.64 to 51.12 on Toys datasets. Those results demonstrated the mechanism of FedCTR.

  2. Training on task heterogenous clients demonstrated the practical importance of FedTCR. 20 clients result also provided evidence for the gain from similar task cooperation. Those evaluations results demonstrate the key practical claim of this paper.

Critical weaknesses
  1. There existed misalignment between proposed FedCTR and key claim of heterogeneity challenges: (1) it claimed that 'This assumption is fragile in MAGs, where node pairs strongly homophilic under one modality may display weak homophily or even heterophily under another', however, there lacks clearly definition of how frequent it happened in the datasets and what's the difference between different clients. (2) what's more, in the implementation, the text and image were using the same weight generated by PageRank, without distinguish different modalities for different nodes.(3) in the experiments, random mask feature or remove the link can only demonstrate the performance when the data is sparse, which can not answer the problem about (4) Based on these points above, the ablation study can only provide evidence to demonstrate the gain from those components, but it can not provide enough evidence for the key claim that it can solve the modality and topology heterogeneity, which largely weaken the soundness of this paper.

  2. Main results can not isolate the contributions from heterogeneous task collaboration and prototypes routing: (1) in figure 2, there lacks of w/o CR on 20 client heterogenous client results, so it remains unclear whether the gain came from larger collaboration(more datapoints/domain knowledge) or from Prototype routing (2) each task only has one dataset, there is no enough evidence to distinguish the different tasks or datasets (3) in conclusion, it can provide evidence from expansion of collaboration range, however, it can not demonstrate the contribution of prototypes routing or the uplift over simple federated pretrain.

Questions for authors
  1. Can the author provide the quantitative analysis of differences in homophily between the textual and visual modalities in the evaluated graphs. What's more, can the same PageRank deal with this kind of inconsistence? If so, is there any related ablation study can support this conclusion?

  2. It is better to provide the result of w/o CR on 20 client heterogenous client under the same data and training budget. How much is the uplift from prototype routing?

  3. Can author explain the difference between similar method such as FedGAMMA.(Li, Xunkai. "Toward Federated Multimodal Graph Foundation Models: A Topology-Aware Multimodal Alignment Framework." arXiv preprint arXiv:2607.15687 (2026).), in terms of problem setting, pretrained method and so on.

3 comments

Thank you for these constructive comments. We address your concerns as follows:


W1. Heterogeneity and shared PageRank weights (Q1). Thank you for your constructive feedback. To quantify cross-modal heterogeneity, we measured edge-wise semantic consistency suv(m)=cos⁡(xu(m),xv(m))s_{uv}^{(m)}=\cos(x_u^{(m)},x_v^{(m)}) using raw text/image features on the original graphs. We report each modality's mean, the cross-modal Spearman correlation ρ\rho, and directional strong–weak conflict: C=12[P(sI≤q25I∣sT≥q75T)+P(sT≤q25T∣sI≥q75I)],C=\frac{1}{2}\left[P(s^I\le q_{25}^I\mid s^T\ge q_{75}^T)+P(s^T\le q_{25}^T\mid s^I\ge q_{75}^I)\right], with quartiles computed within each dataset/client.

DatasetText meanImage meanρ\rhoConflict CC (%)
BiliDance0.7200.766-0.00725.3
RedditS0.6960.8310.06722.4
SemArt0.6380.7740.09822.3

Across all eight datasets, ρ\rho ranges from −0.007 to 0.432 and CC from 9.5% to 25.3%; the three examples above show near-zero rank correlation. BiliDance also varies across five clients (ρ∈[−0.064,0.057]\rho\in[-0.064,0.057], C∈[21.2%,30.7%]C\in[21.2\%,30.7\%], intra-client edges). The independence reference is C=25%C=25\%. Here CC measures directional feature-consistency conflicts on graph edges, complementing label-based homophily. These measurements demonstrate that cross-modal heterogeneity is real and prevalent across the evaluated datasets and clients.

Shared PageRank summarizes the adjacency shared by text and image, while routing and contrastive objectives address cross-modal discrepancies. We compared it with uniform and modality-specific semantic-graph weighting under the same protocol:

VariantMovies ACCRedditS AUCToys R@1
Uniform53.69$\pm$0.3096.89$\pm$0.2152.45$\pm$0.42
Shared PageRank54.27$\pm$0.2597.84$\pm$0.1853.64$\pm$0.36
Modality-specific PageRank54.08$\pm$0.2997.91$\pm$0.2353.51$\pm$0.40

Shared PageRank outperforms Uniform throughout; modality-specific weighting adds no consistent gain. Figure 3 will be described as a sparsity-robustness study.


W2. Isolating routing and task effects (Q2). Thank you for your constructive feedback. To determine whether routing contributes beyond expanded collaboration, we compared FedTCR (Full) and FedTCR (w/o CR) in the 20-client setting using the same data, pre-training budget, and fine-tuning. The latter removes routing and client-level loss.

VariantMovies ACCRedditS AUCToys R@1Flickr30k R-LSemArt C-S
Homo Full (4 clients)53.82$\pm$0.2896.85$\pm$0.2252.62$\pm$0.4126.21$\pm$0.2368.33$\pm$0.27
Hete (w/o CR)53.67$\pm$0.3496.42$\pm$0.2552.21$\pm$0.4626.55$\pm$0.2768.71$\pm$0.31
Hete (Random CR)53.92$\pm$0.3796.88$\pm$0.2452.86$\pm$0.4326.77$\pm$0.3069.24$\pm$0.28
Hete Full54.41$\pm$0.2697.72$\pm$0.1853.91$\pm$0.3527.47$\pm$0.1970.33$\pm$0.24

Hete Full exceeds Hete (w/o CR) by +0.74/+1.30/+1.70/+0.92/+1.62 points, showing the combined contribution of routing and client-level loss. Against Hete (Random CR), which retains this loss, Hete Full gains +0.49/+0.84/+1.05/+0.70/+1.09 points, isolating the benefit of similarity-based routing.

To distinguish task composition from dataset diversity, we compared same- and mixed-task federations on the same Movies graph and 12 partitions, using auxiliary downstream heads during pre-training. Same-task federations are trained separately for NC, LP, and MR; the mixed-task federation combines these objectives. Collaboration scale and budget remain unchanged:

FederationNC ACCLP AUCMR R@1
Same-task53.68$\pm$0.3294.86$\pm$0.2751.74$\pm$0.43
Mixed-task54.02$\pm$0.2995.31$\pm$0.2452.83$\pm$0.37

Since only task composition differs, this comparison separates its effect from dataset diversity; all three metrics favor mixed-task pre-training. We will scope Figure 2 to collaboration across task/dataset pairs.


Q3. Differences from FedGAMMA. Thank you for your constructive feedback. We will cite FedGAMMA as concurrent work. FedGAMMA focuses on semantic and structural alignment; our work jointly formulates task, modality, and topology heterogeneity:

  • different client objectives require task-agnostic pre-training and isolated fine-tuning;
  • modality quality/domain discrepancies require cross-client, cross-modal prototype routing;
  • divergent text/visual edge consistency motivates topology-informed prototypes, and the shared-PageRank ablation evaluates the utility of this design.

Our novelty and uniqueness lie in jointly navigating task, modality, and topology heterogeneity: FedGAMMA studies how to align and aggregate knowledge, whereas FedTCR determines what to share under these disagreements. Its optimal-transport semantic alignment and synchronized prompts contrast with our contrastive references and local adaptation.

Thank you for additional results, which addressed my main concerns. Can the author further make it more clear about the setting of auxiliary loss in the table same- and mixed-task federations on the same Movies graph and 12 partitions ? How does the using auxiliary downstream heads during pre-training work? If the auxiliary loss will update the shared encoder, which will make it different from the original task-agnostic objective, can author explain how this modification can support FedTCR?

Thank you for your follow-up. Here we further clarify the auxiliary objective, the task assignments, and the specific role of this experiment.


Auxiliary objective. The auxiliary heads are used only in this supplementary experiment and are not part of FedTCR’s original pre-training. Let LTCRL_{\mathrm{TCR}} denote the original objective in Eq. (11). A client kk with assigned task tkt_k optimizes

\qquad L_{\mathrm{aux},k}=\ell_{t_k}\left(g_{t_k}(R_k),y_k\right),$$ where $g_{t_k}$ is a task-specific head, $\ell_{t_k}$ is the corresponding NC, LP, or MR loss, and $\eta$ is fixed across clients and settings. Following Section 4.2, $R_k=H_k$ for NC/LP and the modality-specific embeddings for MR. The auxiliary gradient updates the local shared encoder together with $L_{\mathrm{TCR}}$. The auxiliary heads remain client-local; only the shared encoder is aggregated, and prototype routing remains unchanged. After pre-training, all settings undergo the same isolated task-specific fine-tuning and evaluation as the original method. --- **Task assignments.** All settings use the same Movies graph and 12 fixed partitions: |Setting|Task assignment| |-|-| |Same-task|Three separate federations with all 12 clients assigned NC, LP, or MR, respectively| |Mixed-task|One federation with 4 NC, 4 LP, and 4 MR clients| For each target task, evaluation uses the same designated clients and held-out split in both settings. The graph, partitions, collaboration scale, number of training rounds, and auxiliary-loss weight remain unchanged; only the distribution of client objectives differs. --- **Why auxiliary supervision is needed.** Under task-agnostic pre-training, task identities appear only during subsequent fine-tuning. If we changed same-task assignments to mixed-task assignments without auxiliary supervision, the pre-training computation would remain identical. Auxiliary heads therefore provide the task identities needed to examine task composition while holding dataset diversity fixed. --- **How this supports FedTCR.** This experiment is not intended to equate the auxiliary objective with FedTCR’s original task-agnostic objective. Its role is to isolate a different question: whether mixing client objectives can benefit a shared encoder when all clients use the same graph. The positive results on NC, LP, and MR support this task-heterogeneity motivation. The effectiveness of the original task-agnostic framework is evaluated separately by Figure 2, while the new FedTCR (w/o CR) and Random CR comparisons isolate the routing contribution in that original setting. --- Thus, the supplementary table establishes the benefit of mixed-task collaboration without dataset diversity as a confound; the original FedTCR experiments establish that the proposed task-agnostic framework can realize such collaboration without requiring auxiliary task supervision. We will revise the table caption and surrounding text to make this distinction explicit.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.