- 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
-
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.
-
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.
-
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
-
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).
-
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.
-
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.
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):
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:
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, 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.