acceptodds
Under review as a conference paper at ICLR 2027

FedGReS: Federated Generative Recommendation via Global Semantic-ID Relation Sharing

Abstract

Generative recommendation predicts next items by autoregressively generating their hierarchical Semantic IDs (SIDs). We study this paradigm in cross-silo federated learning, where user-item interaction sequences remain local to each client. Learning sequence patterns for next-item prediction is challenging when exclusive-user preferences vary across clients and shared users' interaction histories are fragmented across them. To address these challenges, we propose FedGReS, which complements parameter aggregation with first-level SID graphs built from pooled transition and co-occurrence counts. Distinct item pairs mapped to the same pair of first-level SID codewords contribute to shared counts, making cross-client relational evidence available to local sequence learning. The resulting graphs condition first-level SID representations of historical items through graph propagation and gated residual injection. The graph-conditioning mechanism and the generative model are jointly trained under the full-SID generation objective. FedGReS exchanges neither user identifiers nor interaction sequences. Across three Amazon Reviews 2023 datasets, FedGReS achieves the highest mean performance in all 12 dataset-metric settings, with relative gains of up to 6.98% over the strongest baseline.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Generative Recommendation, Federated Recommendation, Semantic IDs

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_FtxyOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a federated generative recommendation model using hierarchical Semantic IDs, namely FedGReS. Specifically, it builds shared relation graphs from SID transition and co-occurrence statistics collected on local clients. Two graph-attention modules then refine the SID representations for a TIGER-based recommender. Experiments are conducted on three Amazon Reviews 2023 datasets, with comparisons against federated recommendation baselines.

Critical strengths
  1. The idea of sharing coarse Semantic ID relations is reasonable.
  2. The experimental evaluation is fairly comprehensive.
Critical weaknesses
  1. Although different client partitions are evaluated, it is unclear how much of the improvement comes from addressing history fragmentation rather than general relation sharing.
  2. The related-work section explicitly discusses FedCGR (Liu et al., 2026), which also studies SID-based federated generative recommendation, but FedCGR is not included in Table 1.
  3. Since secure aggregation is optional, the server may access client-level relation statistics. It is unclear whether these shared statistics could reveal sensitive information. Thus, the privacy protection may need further clarification.
  4. The evaluation under different client heterogeneity settings is limited to the Video dataset.
Questions for authors
  1. Could the authors provide more direct evidence that FedGReS helps address fragmented shared-user histories?
  2. Would it be possible to add FedCGR (Liu et al., 2026) as one of the baseline methods?
  3. Could the authors clarify the privacy implications of sharing client-level relation statistics?
  4. Could the authors extend the client heterogeneity experiments to Software and Baby datasets?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Ftxy [1/2]

Dear Reviewer Ftxy,

We sincerely thank the reviewer for your time and insightful feedback. We are grateful for the positive assessment of our work and the constructive comments, which help us significantly improve our paper. All comments have been carefully addressed, and the corresponding revisions have been updated in the revised paper with major changes highlighted in purple for clarity. We hope our revisions address all the concerns and bring the paper up to acceptance.

Below, we provide point-by-point responses to each comment, along with additional explanations and experimental results where necessary.

  • Weakness 1 and Question 1. Fragmented shared-user histories

Response to Weakness 1 and Question 1. Thank you for raising this point. We agree that the benefit to fragmented shared-user histories should be examined more directly. We have added a targeted analysis in Section 5.3 and its evaluation protocol in Appendix D.2 of the revised manuscript.

We reused the original three-client partitions and model checkpoints from the main comparison in Section 5.2 on Video, Software, and Baby. The analysis focuses on shared-user predictions for which local relational evidence is weak and additional evidence is available from other clients. For a test record ee, let aea_e and beb_e denote the first-level SIDs of the last input-history item and the target item, respectively. For each graph type g∈{tr,co}g\in\{\mathrm{tr},\mathrm{co}\}, Lg(e)L_g(e) is the local training count of this relation, Rg(e)R_g(e) is its summed training count on the other clients, and AaebegA^g_{a_e b_e} is its weight in the deployed pooled graph. We define

B = \left\\{e \in \mathcal{D}_{\mathrm{shared}} : \exists g \in \\{\mathrm{tr},\mathrm{co}\\},\\; L_g(e) \leq 20 \land R_g(e) > L_g(e) \land A^g_{a_e b_e} > 0\right\\}.

where Dshared\mathcal{D}_{\mathrm{shared}} contains the client-local test records of shared users. All denotes the full test set, including BB.

Test records (All / B): Video 16,625 / 4,497; Software 32,514 / 5,722; Baby 18,464 / 3,761.

DatasetSubsetR@5R@10
FedAvgFedGReSgainFedAvgFedGReSgain
VideoAll0.03300.0355+7.58%0.05300.0566+6.79%
B0.05160.0576+11.63%0.08090.0901+11.37%
SoftwareAll0.06980.0745+6.73%0.10890.1134+4.13%
B0.08840.0952+7.69%0.13880.1456+4.90%
BabyAll0.01720.0184+6.98%0.02610.0267+2.30%
B0.04570.0497+8.75%0.06510.0686+5.38%

Across all three datasets and both cutoffs, FedGReS achieves larger relative gains on BB than on All. The absolute Recall gains are also larger on BB in every comparison. These results support FedGReS's ability to mitigate the effects of fragmented shared-user histories by supplementing weak local relational evidence with transition and co-occurrence patterns from other clients.

  • Weakness 2 and Question 2. Comparison with FedCGR

Response to Weakness 2 and Question 2. Thank you for this suggestion. We have added FedCGR as one of the baseline methods. FedCGR's original backbone performed better during tuning, so we retained it. We follow the original FedCGR settings for the generator and SASRec architectures, CF adapter, private experts, expert dimensions and routing, auxiliary loss weights, contrastive temperature, prototype EMA, and generator batch size.

Tuned hyperparameters:

  • Shared experts: 3.
  • Generator dropout: 0.35 for Video/Baby and 0.25 for Software.
  • Generator learning rate: 10⁻³.
  • CF pretraining: 60 epochs.
  • Proximal coefficient μ: 10⁻⁵.
  • Label smoothing: 0.05.

Main comparison. The table below compares FedCGR and FedGReS on all three datasets.

MethodVideoSoftwareBaby
R@5N@5R@10N@10R@5N@5R@10N@10R@5N@5R@10N@10
FedCGR0.03320.02080.05420.02750.07220.04580.11190.05950.01710.01090.02620.0139
FedGReS0.03550.02260.05660.02940.07450.04830.11340.06090.01840.01210.02670.0148

FedCGR is competitive with the existing baselines. Nevertheless, FedGReS outperforms FedCGR on all four metrics across all three datasets, with relative Recall@5 improvements of 6.93%, 3.19%, and 7.60% on Video, Software, and Baby, respectively.

Data heterogeneity and federation size. We also evaluated FedCGR on Video under varying data heterogeneity and federation sizes. We have updated Sections 5.2 and 5.6 of the revised manuscript to include these additional comparisons. FedGReS consistently outperforms FedCGR on all four metrics across the tested settings.

AuthorsOpenReview ↗

Response to Reviewer Ftxy [2/2]

  • Weakness 3 and Question 3. Privacy implications

Response to Weakness 3 and Question 3. Thank you for requesting clarification. We have revised the “Privacy and communication scope” paragraph in Section 4.2 and added Appendix B. FedGReS aims to protect user identities, specific user–item associations, and records identifying which item pairs occur in a local sequence. Coarse first-level SID relations are intentionally shared as collaborative knowledge and fall outside the information we aim to conceal.

Even when the server receives raw client-level counts, the matrices record first-level SID transitions and co-occurrences aggregated across users, without identifying the contributing item pairs, users, or sequences. Replacing items with locally available alternatives sharing the same first-level codewords, while preserving sequence positions and boundaries, leaves both count matrices unchanged. Reassigning complete histories among users within the same client likewise preserves the counts. A specific user–item or item-pair fact is therefore not uniquely determined when such alternatives remain compatible with the observer's knowledge and differ on that fact. This ambiguity is inherent to the count representation and does not depend on secure aggregation.

Without secure aggregation, the server can still observe each client's coarse relation patterns and exact counts. We explicitly identify this disclosure as a limitation in Appendix B. For applications requiring confidentiality of these patterns, future work will investigate controlled count perturbation and its privacy–utility trade-off.

  • Weakness 4 and Question 4. Heterogeneity experiments on Software and Baby datasets

Response to Weakness 4 and Question 4. Thank you for this suggestion. We agree that extending the heterogeneity evaluation beyond Video strengthens the empirical assessment. We conducted additional experiments on Software and Baby, prioritizing comparison with FedProx, a strong baseline in the Video study that explicitly addresses heterogeneous federated optimization. Experiments with the remaining baselines are ongoing, and we will supplement the revised manuscript with the complete comparisons once they finish.

We follow the same data-partitioning and training protocols as the Video study, using the hyperparameter-selection procedure described in Appendix C.4.

Dataset(N,ρex,γ)(N,\rho_{\mathrm{ex}},\gamma)MethodR@5N@5R@10N@10
Software(3,0.6,10)(3,0.6,10)FedProx0.07390.04770.11210.0600
(3,0.6,10)(3,0.6,10)FedGReS0.07340.04740.11180.0599
(3,0.9,10)(3,0.9,10)FedProx0.07480.04910.11430.0618
(3,0.9,10)(3,0.9,10)FedGReS0.07780.05060.11840.0637
(3,0.3,1)(3,0.3,1)FedProx0.07160.04620.11130.0590
(3,0.3,1)(3,0.3,1)FedGReS0.07690.05090.11480.0631
(3,0.3,3)(3,0.3,3)FedProx0.07070.04500.11090.0580
(3,0.3,3)(3,0.3,3)FedGReS0.07450.04750.11520.0606
(5,0.3,10)(5,0.3,10)FedProx0.06660.04320.10200.0547
(5,0.3,10)(5,0.3,10)FedGReS0.07030.04610.10540.0575
(7,0.3,10)(7,0.3,10)FedProx0.05630.03650.08990.0473
(7,0.3,10)(7,0.3,10)FedGReS0.05940.03840.09130.0488
Baby(3,0.6,10)(3,0.6,10)FedProx0.01630.01070.02600.0138
(3,0.6,10)(3,0.6,10)FedGReS0.01810.01180.02660.0145
(3,0.9,10)(3,0.9,10)FedProx0.01780.01140.02770.0146
(3,0.9,10)(3,0.9,10)FedGReS0.01790.01150.02850.0149
(3,0.3,1)(3,0.3,1)FedProx0.01870.01170.02730.0145
(3,0.3,1)(3,0.3,1)FedGReS0.01870.01170.02780.0147
(3,0.3,3)(3,0.3,3)FedProx0.01640.01080.02440.0135
(3,0.3,3)(3,0.3,3)FedGReS0.01650.01040.02460.0131
(5,0.3,10)(5,0.3,10)FedProx0.01470.00880.02220.0112
(5,0.3,10)(5,0.3,10)FedGReS0.01520.00940.02200.0117
(7,0.3,10)(7,0.3,10)FedProx0.01520.00970.02120.0117
(7,0.3,10)(7,0.3,10)FedGReS0.01610.01000.02150.0117

Across the 48 dataset–configuration–metric comparisons, FedGReS achieves higher scores in 38, matches FedProx in 3, and obtains lower scores in 7. The 7 deficits range from 0.0001 to 0.0005 in absolute score. These results show that FedGReS remains competitive beyond Video and support its effectiveness across the tested datasets, client distributions, and federation sizes.

I appreciate the authors' substantial efforts in revising the manuscript. The new subgroup analysis provides more direct evidence of the benefits of FedGReS for fragmented shared-user histories, and the inclusion of FedCGR makes the baseline comparison more complete. The revised privacy discussion also clarifies the scope and limitations of sharing coarse-grained relation statistics. Additional heterogeneity results on Software and Baby are also helpful.

Overall, I appreciate the authors' efforts and the improvements made in the revised manuscript. I have therefore decided to increase my rating from 2 to 3.

AuthorsOpenReview ↗

Response to Reviewer Ftxy

Dear Reviewer Ftxy,

We sincerely thank you for your valuable feedback and for updating your rating. We are glad that our revisions addressed your concerns.

Reviewer_zm93OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FedGReS, a federated generative recommendation framework that shares behavioral relations through first-level Semantic IDs (SIDs). To address heterogeneous preferences among client-exclusive users and fragmented histories of shared users, FedGReS aggregates client-local transition and co-occurrence statistics into global SID-level relation graphs. These graphs are encoded using dual graph attention networks and incorporated into the generative recommender through gated residual fusion. Experiments on three Amazon Reviews 2023 datasets show consistent average improvements over federated recommendation baselines. The paper also provides component ablations, design comparisons, sensitivity analyses, and graph- and representation-level analyses.

Critical strengths

1. A useful approach to cross-client relational knowledge sharing.

Using first-level SIDs as shared aggregation units allows behavioral relations from different item pairs and clients to be pooled without explicit user matching or raw interaction exchange. This provides a compact, semantically aligned representation for transferring relational knowledge beyond conventional parameter aggregation. The use of coarse SID groups for relation sharing while retaining complete SIDs for item prediction is a sensible design choice.

2. A reasonably comprehensive empirical evaluation.

The paper evaluates FedGReS against eight federated baselines across three datasets, reporting consistent improvements in average recommendation performance. Beyond the main comparisons, the component ablations and design alternatives, including different SID relation groupings and graph-conditioning strategies, provide useful evidence for the proposed architectural choices. The additional relation case studies and representation analyses further support the technical evaluation.

Critical weaknesses

Insufficient experimental validation of the central research challenges.

The main concern is the disconnect between the stated research challenges and the experimental evaluation. The paper identifies heterogeneous preferences among client-exclusive users and fragmented histories of shared users as its two principal challenges. However, the experiments do not adequately isolate either challenge or establish how the proposed mechanism specifically addresses them.

First, varying the exclusive-user ratio does not directly control the degree of preference divergence across clients. The partitioning protocol distributes exclusive users across clients without explicitly constructing or measuring differences in their preference distributions. Consequently, it remains unclear whether global SID relation sharing mitigates preference heterogeneity or could introduce negative transfer when client-specific behavioral patterns conflict.

Second, varying the number of clients and the Dirichlet allocation parameter provides some relevant evidence regarding history fragmentation, but these changes also affect local sequence lengths, data availability, and potentially the evaluation populations. More importantly, the experiments do not establish whether the observed improvements specifically arise from relational information that is missing locally but available from other clients.

Although the w/o Agg ablation supports the benefit of global relation sharing, the graph-fidelity and representation analyses do not sufficiently explain how this benefit relates to the stated challenges. Closer agreement with a centralized reference graph and improved representation alignment do not necessarily demonstrate that missing sequential information is effectively compensated for. Thus, while the results support the utility of pooled SID relations, the evidence remains insufficient to substantiate the paper's central problem-specific claims.

Questions for authors
  1. Can the authors provide evidence that global SID relation sharing specifically mitigates inter-client preference heterogeneity? For example, how does its benefit change when preference divergence across clients is explicitly controlled or measured?

  2. Can the authors show whether the improvements from global relation sharing are concentrated in cases where relevant SID relations are absent locally but observed by other clients? Such an analysis would help establish whether the proposed mechanism actually compensates for fragmented interaction histories.

2 comments
AuthorsOpenReview ↗

Response to Reviewer zm93 [1/2]

Dear Reviewer zm93,

We sincerely thank the reviewer for your time and insightful feedback. We are grateful for the positive assessment of our work and the constructive comments, which help us significantly improve our paper. All comments have been carefully addressed, and the corresponding revisions have been updated in the revised paper with major changes highlighted in purple for clarity. We hope our revisions address all the concerns and bring the paper up to acceptance.

Below, we provide point-by-point responses to each comment, along with additional explanations and experimental results where necessary.

  • Weakness 1. Insufficient experimental validation of the central research challenges

Response to Weakness 1. Thank you for highlighting this gap. We agree that the two research challenges require separate, targeted evaluation. We have conducted two additional evaluations and incorporated their results into Section 5.3, with the protocols detailed in Appendices D.1 and D.2 of the revised manuscript.

Across all three datasets, FedGReS achieves larger relative Recall gains over FedAvg under stronger measured preference divergence among exclusive users (Q1). It also achieves larger relative Recall gains on shared-user predictions with weak local but stronger cross-client relation support than on the full test set (Q2). Together, these findings provide targeted empirical support for FedGReS's ability to mitigate both preference heterogeneity and fragmented shared-user histories. The experimental settings, results, and analyses are provided in our responses to Q1 and Q2 below.

  • Question 1. Preference heterogeneity

Response to Question 1. Thank you for this suggestion. We agree that preference divergence should be explicitly controlled and measured. We conducted additional experiments with three clients and only exclusive users (ρex=1\rho_{\mathrm{ex}}=1). We have added these experiments in Section 5.3 and their evaluation protocol in Appendix D.1 of the revised manuscript.

Each user's training, validation, and test records belonged to a single client, separating preference heterogeneity from shared-user history fragmentation. We measured preference divergence using the average pairwise Jensen–Shannon divergence between clients' first-level SID frequency distributions:

H=\frac{1}{3}\sum_{1\leq i<j\leq3}\mathrm{JS}(p_i,p_j).$$ Here, $c_k(s)$ counts interactions with first-level SID $s$ in client $k$'s complete original training prefixes, and JS uses base-2 logarithms. Thus, $H$ measures differences in the marginal distributions of semantic groups across clients. For each dataset, we constructed lower- and higher-$H$ partitions by exchanging complete users between clients, using training statistics only. We recomputed $H$ on the resulting partitions and compared FedGReS with FedAvg under each setting. The results are shown below. | Dataset | H | | R@5 | | | R@10 | | |---|---:|---:|---:|---:|---:|---:|---:| | | | FedAvg | FedGReS | gain | FedAvg | FedGReS | gain | | Video | 0.10 | 0.0289 | 0.0304 | +5.19% | 0.0456 | 0.0503 | +10.31% | | | 0.30 | 0.0258 | 0.0308 | **+19.38%** | 0.0435 | 0.0487 | **+11.95%** | | Software | 0.10 | 0.0726 | 0.0759 | +4.55% | 0.1089 | 0.1132 | +3.95% | | | 0.30 | 0.0709 | 0.0746 | **+5.22%** | 0.1088 | 0.1140 | **+4.78%** | | Baby | 0.03 | 0.0151 | 0.0163 | +7.95% | 0.0232 | 0.0247 | +6.47% | | | 0.09 | 0.0136 | 0.0151 | **+11.03%** | 0.0223 | 0.0241 | **+8.07%** | FedGReS outperforms FedAvg at both heterogeneity levels on all three datasets. Its relative gains in both Recall metrics are larger at higher H. These controlled results support FedGReS's ability to mitigate inter-client preference heterogeneity, with greater benefits under stronger measured preference divergence.
AuthorsOpenReview ↗

Response to Reviewer zm93 [2/2]

  • Question 2. Fragmented interaction histories

Response to Question 2. Thank you for this suggestion. We agree that the benefit of cross-client relation sharing should be examined directly in cases with limited local relational evidence. We have added a targeted analysis in Section 5.3 and its evaluation protocol in Appendix D.2 of the revised manuscript.

Using the same data partitions and model checkpoints as the main comparison in Section 5.2, we compared FedGReS and FedAvg on shared-user predictions with weak local relation support and stronger support from other clients.

For each test record ee, let aea_e be the first-level SID of the last item in its input history and beb_e that of its target item. For each graph type g∈{tr,co}g\in\{\mathrm{tr},\mathrm{co}\} , let Lg(e)L_g(e) denote the count of this relation on the record's client and Rg(e)R_g(e) its summed count on the other clients. Let AaebegA^g_{a_e b_e} denote the corresponding weight in the deployed pooled graph after pruning and normalization. We define

B = \left\\{e \in \mathcal{D}_{\mathrm{shared}} : \exists g \in \\{\mathrm{tr},\mathrm{co}\\},\\; L_g(e) \leq 20 \land R_g(e) > L_g(e) \land A^g_{a_e b_e} > 0\right\\}.

Both methods were evaluated on the same fixed BB. All denotes the full test set, including BB.

Test records (All / B): Video 16,625 / 4,497; Software 32,514 / 5,722; Baby 18,464 / 3,761.

DatasetSubsetR@5R@10
FedAvgFedGReSgainFedAvgFedGReSgain
VideoAll0.03300.0355+7.58%0.05300.0566+6.79%
B0.05160.0576+11.63%0.08090.0901+11.37%
SoftwareAll0.06980.0745+6.73%0.10890.1134+4.13%
B0.08840.0952+7.69%0.13880.1456+4.90%
BabyAll0.01720.0184+6.98%0.02610.0267+2.30%
B0.04570.0497+8.75%0.06510.0686+5.38%

Across all three datasets and both cutoffs, FedGReS achieves larger relative gains on BB than on the full test set. These results show that FedGReS particularly benefits shared-user predictions with limited local relational evidence and stronger evidence available through the pooled graph. They provide targeted empirical support for alleviating fragmented shared-user histories by supplementing local sequence modeling with cross-client SID relations.

Reviewer_TEBdOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedGReS, a cross-silo federated generative recommendation framework that complements standard parameter aggregation (FedAvg) with global semantic-ID relation sharing. To address the challenges of heterogeneous exclusive-user preferences and fragmented shared-user histories across clients, the method pools first-level Semantic ID (SID) transition and co-occurrence counts into fixed sparse graphs on the server without exchanging raw sequences or user identifiers. During local training, dual Graph Attention Networks (GATs) propagate over these pooled graphs to refine first-level SID embeddings via gated residual injection, which are then used by an autoregressive Transformer for full-SID generation. Experiments on three Amazon Reviews 2023 datasets demonstrate that FedGReS consistently outperforms eight federated baselines, achieving relative gains of up to 6.98%.

Critical strengths

Strengths:

1) The motivation is well-grounded; addressing fragmented shared-user histories and heterogeneous preferences in cross-silo federated generative recommendation is a practical and underexplored problem.

2) Pooling first-level SID transition and co-occurrence counts is a clever design choice that allows cross-client relational evidence sharing while naturally preserving privacy and avoiding complex cross-client user matching.

3) The paper provides thorough empirical analyses beyond main results, including graph fidelity comparisons against centralized references, representation alignment margins, and detailed ablation studies validating individual components.

4) The communication overhead is highly efficient, as the relation graphs are constructed and broadcast only once before training, adding negligible cost compared to per-round model synchronization.

Critical weaknesses

Weaknesses:

1) The absolute performance across all methods is critically low (e.g., Recall@5 around 0.018–0.035), suggesting the task formulation or data partitioning is unrealistically difficult. A 6.98% relative improvement on such marginal baselines yields negligible practical utility, making it unclear whether the proposed complexity actually solves a viable recommendation problem or merely slightly mitigates a flawed setup.

2) Aggregating relations exclusively at the first-level SID introduces severe information loss. Because the first codeword represents the coarsest semantic cluster, pooling transitions at this level conflates distinct fine-grained item dynamics. For instance, users transitioning between two broad categories does not guarantee they transition between specific items within those categories, potentially injecting noisy structural priors into the local sequence modeling.

3) The relation graphs are constructed once and remain static throughout the entire federated optimization process. As the model learns and client-local sequence distributions shift during training, the initial count-based graphs may become stale or misaligned with the evolving embedding space, yet the framework lacks any mechanism to iteratively update or refine these topologies.

4) While the authors claim compatibility with secure aggregation, broadcasting the exact K×K transition and co-occurrence matrices back to clients can still leak sensitive behavioral patterns. An adversary observing the global graph could potentially infer dominant sequential behaviors or the presence of specific item-category transitions within a silo, undermining the strict privacy guarantees expected in cross-silo settings.

Questions for authors

Questions:

1) Given the extremely low absolute metrics (e.g., R@5 < 0.04), could the authors clarify if this stems from the artificial Dirichlet fragmentation of shared-user sequences, and how the method performs when sequence fragmentation is less aggressive?

2) How sensitive is the overall performance to the codebook size K at the first level? If K is too small, the graph becomes overly dense and noisy; if too large, it remains sparse and fails to capture cross-client overlaps. Is there an optimal trade-off analyzed?

3) Since the graphs are static, did the authors experiment with periodically updating the pooled relation graphs every few communication rounds using the latest model predictions, rather than relying solely on the initial training-prefix counts?

6 comments
AuthorsOpenReview ↗

Response to Reviewer TEBd [1/2]

Dear Reviewer TEBd,

We sincerely thank the reviewer for your time and insightful feedback. We are grateful for the positive assessment of our work and the constructive comments, which help us significantly improve our paper. All comments have been carefully addressed, and the corresponding revisions have been updated in the revised paper with major changes highlighted in purple for clarity. We hope our revisions address all the concerns and bring the paper up to acceptance.

Below, we provide point-by-point responses to each comment, along with additional explanations and experimental results where necessary.

  • Weakness 1 and Question 1. Low absolute metrics

Response to Weakness 1 and Question 1. Thank you for raising these concerns. First, FedGReS achieves a Recall@5 of 0.0745 on Software. Second, low Recall@5 values also occur in established centralized generative recommendation benchmarks: TIGER [1] (Table 1) reports 0.0264 on Sports, while GRID [2] (Table 1) also reports 0.0236 on Sports. These results contextualize the numerical scale. Our implementation builds on GRID and follows its model configuration, allowing us to study federated relation sharing on an established generative recommendation framework.

To directly address Q1, we conducted additional FedGReS experiments with less aggressive fragmentation. The original setting uses N=3N=3, ρex=0.3\rho_{\mathrm{ex}}=0.3, and γ=10\gamma=10, with most shared users' histories distributed across three clients. In the new setting, we redistribute each shared user's history across only two of the three clients, yielding longer local subsequences.

DatasetSettingRecall@5NDCG@5Recall@10NDCG@10
VideoOriginal0.03550.02260.05660.0294
Less fragmented0.03580.02250.05740.0291
SoftwareOriginal0.07450.04830.11340.0609
Less fragmented0.07550.04910.11600.0625
BabyOriginal0.01840.01210.02670.0148
Less fragmented0.01860.01210.02800.0150

Recall@5 increases by only 0.0003, 0.0010, and 0.0002 on Video, Software, and Baby, respectively. These results show that this scale persists under less aggressive fragmentation and is not specific to the original partition.

  • Weakness 2. Information loss of first-level SID

Response to Weakness 2. Thank you for raising this concern. Sharing first-level SID relations adds supplementary information without discarding fine-grained input information from local histories. The Transformer still receives complete historical SIDs and generates complete target SIDs, learning coarse-to-fine dependencies. Aligned semantic groups pool sparse cross-client observations without imposing item-specific transition rules, while attention and gated residual fusion augment only first-level history embeddings, preserving the original embedding path.

FedGReS outperforms graph-free FedAvg, w/Rand-Rel with random item groups, and w/o Gate, supporting relation sharing and conditioning, semantic grouping, and gated fusion, respectively (Sections 5.2 and 5.4). Graph conditioning also improves first-level representation margins and relation-retrieval nDCG@10 across all three datasets (Section 5.7 and Appendix G.3). These results support coarse relations for fine-grained recommendation.

  • Weakness 3 and Question 3. Static relation graphs

Response to Weakness 3 and Question 3. Thank you for raising these questions. FedGReS uses a static behavioral prior with dynamic representation learning. With fixed training data and SID mappings, the observed SID relations are independent of embedding coordinates, so embedding updates alone do not make the prior stale. Embeddings, GAT parameters, and gates are jointly optimized for complete-SID generation, and graph-conditioned representations are recomputed at every local optimization step. Consequently, the injected graph representations evolve with the embedding space instead of remaining fixed at initialization.

With fixed graphs, FedGReS outperforms graph-free FedAvg, and graph conditioning improves representation margins and relation-retrieval nDCG@10 across all three datasets. These results support the effectiveness of the static prior.

Regarding Q3, the training data and SID mappings remain fixed, so the observed first-level relation counts do not change during optimization. Periodic prediction-based updates would introduce inferred relations beyond the observed statistics, constituting a separate mechanism for future research.

AuthorsOpenReview ↗

Response to Reviewer TEBd [2/2]

  • Weakness 4. Privacy concern

Response to Weakness 4. Thank you for highlighting the unclear privacy scope. We have revised the “Privacy and communication scope” paragraph in Section 4.2 and added Appendix B. FedGReS aims to protect user identities, specific user–item associations, and records identifying which item pairs occur in a local sequence. Coarse first-level SID relations are intentionally shared as collaborative knowledge and fall outside the information we aim to conceal.

  1. Omission of item and user provenance. Counts omit specific item pairs and user/sequence attribution. Different item histories with identical first-level SID sequences, or different assignments of complete histories to users within a client, yield identical counts. A specific association is not uniquely determined when these alternatives remain compatible with the observer's knowledge and differ on that association.
  2. Restricted access to client counts. When enabled, secure aggregation gives the server only pooled counts, without direct access to individual client matrices.
  3. Reduced graph disclosure. Clients receive selected relations and within-row relative weights after pruning and normalization, without raw counts or row totals.

The published graphs still reveal aggregate coarse patterns. Without secure aggregation, the server also sees client-level coarse patterns and exact counts. These disclosures are outside our protection scope. Future work will investigate controlled count perturbation and its privacy–utility trade-off for settings requiring confidential coarse patterns.

  • Question 2. Sensitivity to Codebook Size K

Response to Question 2. Thank you for suggesting this sensitivity analysis. The codebook size K controls the number of discrete codewords available at each quantization level and influences the semantic granularity of SIDs, making it an important parameter even in centralized generative recommendation. TIGER [1] (Appendix E) reports robust recommendation performance under the tested changes to SID length and codebook size. GRID [2] (Table 3) compares three-level RK-Means with K=128, 256, and 512 on Beauty, obtaining Recall@5 of 0.0412, 0.0422, and 0.0415, respectively. These results provide an empirical basis for the three-level, K=256 configuration. Our implementation builds on GRID and follows its generative backbone configuration; accordingly, we adopted K=256 in the original experiments.

Since FedGReS additionally constructs relation graphs over first-level SIDs, we agree that its sensitivity should also be examined in the federated setting. We conducted additional experiments with K=128 and K=512 on all three datasets. The table below reports performance:

DatasetKRecall@5NDCG@5Recall@10NDCG@10
Video1280.03590.02260.05630.0293
256 (original)0.03550.02260.05660.0294
5120.03470.02210.05610.0287
Software1280.07480.04860.11450.0611
256 (original)0.07450.04830.11340.0609
5120.07520.04870.11430.0612
Baby1280.01850.01200.02690.0147
256 (original)0.01840.01210.02670.0148
5120.01820.01180.02620.0145

The results indicate that FedGReS is robust to codebook size K. They complement the prior SID studies with evidence of stable recommendation performance in FedGReS, supporting flexibility in codebook selection for practical use.

[1] Rajput S, Mehta N, Singh A, et al. "Recommender systems with generative retrieval." NeurIPS 2023.

[2] Ju C M, Collins L, Neves L, et al. "Generative recommendation with semantic ids: a practitioner's handbook." CIKM 2025.

Thanks for the detailed rebuttal. After reading the rebuttal and the reviews from other reviewers, I decide to remain my score due to the limited improvement and applications.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.