acceptodds
Under review as a conference paper at ICLR 2027

Connecting Autonomous Recommender Systems through Federated Semantic Grounding

Abstract

Autonomous recommender systems learn from different item offerings and behavioral feedback, and thus acquire complementary knowledge about items and user preferences. However, such knowledge is deeply embedded in system-specific recommendation spaces, making direct cross-system transfer difficult. Federated recommendation enables decentralized knowledge exchange, but existing methods often rely on entity correspondence or compatible model and representation spaces. LLMs offer a shared semantic basis across systems, but their generic semantics do not capture the recommendation meaning learned from local interactions. We propose Federated Semantic Grounding, which grounds shared semantics in local behavior and uses them to transfer recommendation knowledge across systems. The framework applies this principle at both the item and user levels. At the item level, we use shared semantic concepts as anchors to exchange complementary item semantics across different item spaces. At the user level, we re-express locally learned preferences in another system's item context and coordinate their interpretation using mappings learned from local behavior. Experiments across diverse collaboration scenarios show consistent improvements over independent and federated baselines, and further demonstrate that FedSG remains effective when participating systems differ in model architecture or local recommendation setting. These results support semantic grounding as an effective basis for collaboration among autonomous recommender systems. The code is available at: https://anonymous.4open.science/r/FedSG-683D/.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Cross-platform Recommendation, Federated Learning, Semantic Grounding

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NaPNOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates knowledge collaboration among autonomous recommender systems without known cross-system user or item correspondences and with potentially heterogeneous model architectures. The authors propose Federated Semantic Grounding (FedSG), which uses shared semantic concepts as a collaboration interface and grounds them in local behavioral feedback. At the item level, FedSG selectively complements under-supported local semantics through shared semantic anchors. At the user level, it re-expresses local preferences in another system's recommendation context and coordinates preference mappings across systems. Experiments on three Amazon-based collaboration scenarios demonstrate improvements over independent and federated baselines across multiple recommendation backbones. Ablation studies and sensitivity analyses further examine the contributions of the proposed components and their dependence on coordination settings.

Critical strengths

1. A meaningful approach to knowledge collaboration across autonomous recommenders.

FedSG addresses an important setting where user and item correspondences are unavailable and recommendation models are independently maintained. Its combination of behavior-grounded item semantics and cross-context user preference transfer provides a useful extension to existing semantic-based federated recommendation approaches. This is a substantive contribution beyond conventional model-level collaboration.

2. Broad and generally consistent empirical support.

The experiments cover three collaboration scenarios, multiple recommendation backbones, and heterogeneous model configurations. FedSG achieves consistent performance improvements over representative baselines, while ablation studies support the contributions of both coordination pathways. The sensitivity analyses also clarify that stronger semantic coordination does not necessarily yield better performance. Together, these results provide meaningful support for the proposed framework.

Critical weaknesses

1. Insufficient clarification of cross-system parameter compatibility.

FedSG establishes shared semantic anchors across systems, but aligning semantic concepts does not necessarily make their independently learned representations directly compatible. Specifically, Eq. (5) averages embeddings associated with the same semantic descriptors across systems. However, these embeddings are learned from different local interactions and may have different coordinate meanings, even when they represent the same concept. Similarly, Eq. (9) combines preference mappings learned independently by different systems, which requires compatible parameter structures and corresponding parameter meanings. The paper does not sufficiently explain how these conditions are ensured, particularly when the participating recommenders use heterogeneous backbones (Tables 3 and 7). Since collaboration without model-space alignment is a central motivation of FedSG, clarifying how its shared semantic components remain compatible is important for establishing the technical soundness of the proposed framework.

2. Limited evidence for the effectiveness of cross-context preference re-expression.

A key component of FedSG is re-expressing locally learned user preferences in another system's recommendation context through LLM-generated semantic profiles and pseudo-reviews. However, the current experiments do not sufficiently isolate the contribution of this re-expression process. Table 4 establishes the benefit of preference coordination, while Appendix G.4 examines the number of generated examples, but neither directly distinguishes meaningful cross-context preference transfer from the benefits of additional LLM-generated semantic features. Consequently, it remains unclear whether the improvements arise from preserving transferable user-specific preferences or from general semantic augmentation. This limits the evidence supporting the proposed user-level grounding mechanism, although it does not undermine the overall performance improvements.

Questions for authors

Q1 (Technical clarification). How is parameter compatibility ensured when aggregating independently learned semantic embeddings and preference mappings in Eqs. (5) and (9)? Please clarify the compatibility mechanism, particularly for heterogeneous backbones with different representation dimensions (Tables 3 and 7).

Q2 (Additional evidence). Can the authors further isolate the contribution of cross-context preference re-expression, for example, by comparing it with source-domain preference profiles without re-expression or user-shuffled target-domain pseudo-reviews? Such comparisons would help establish whether the gains reflect transferable user-specific preferences rather than general semantic augmentation.

Q3 (Evaluation protocol clarification). Are the LLM-generated preference profiles and pseudo-reviews constructed exclusively from training interactions, excluding validation and test interactions under the leave-one-out protocol in Appendix D.3?

1 comment
AuthorsOpenReview ↗

Response to Reviewer NaPN

Response to Q1 and Weakness 1: Technical clarification.

We appreciate the reviewer's insightful question. We agree that shared semantic concepts alone do not guarantee parameter compatibility, and would like to clarify the design underlying our federated coordination.

First, the heterogeneity addressed by FedSG concerns the underlying recommendation models, while the semantic components involved in federated coordination adopt compatible architectures and parameter dimensions across systems. This distinction provides the basis for the two coordination mechanisms.

For item-level coordination, all systems use the same semantic schema and initialize corresponding descriptor embeddings from identical pretrained representations. Thus, Eq. (5) aggregates embeddings with consistent semantic definitions and a common initialization, rather than independently constructed representation spaces.

For user-level coordination, all systems use the same pretrained text encoder, ensuring a shared input representation space, while the preference mappings adopt compatible architectures and parameter dimensions. This provides a common interface for the parameter coordination in Eq. (9), independent of the underlying recommendation models.

We thank the reviewer for highlighting this important distinction, which helps us clarify the compatibility assumptions and better establish the technical soundness of the proposed framework.

Response to Q2 and Weakness 2: Additional evidence

We thank the reviewer for suggesting a more targeted comparison to distinguish cross-context preference transfer from general LLM-based semantic augmentation.

To address this, we introduce a new ablation, "w/o Cross-context Re-expression," which directly uses each user's source-domain preference profile without re-expression, rather than using the generated target-domain pseudo-reviews. All other components and settings, including cross-system preference mapping coordination and model capacity, remain unchanged. This controlled comparison examines whether the gains arise from cross-context preference transfer rather than merely incorporating additional LLM-generated semantic information.

For ease of comparison, we present the new results (in red) alongside the previously reported ablation results from Table 4. FedSG consistently outperforms the new variant across all three scenarios and evaluation metrics, with particularly pronounced gains in Scenario 3. These results further support the effectiveness of cross-context preference transfer beyond general semantic augmentation.

VariantScenario 1 MRRScenario 1 H@10Scenario 1 N@10Scenario 2 MRRScenario 2 H@10Scenario 2 N@10Scenario 3 MRRScenario 3 H@10Scenario 3 N@10
w/o Item Coordination0.19480.37780.22070.19930.36220.22180.18500.33240.2024
w/o Preference Coordination0.19050.37030.21520.21130.38380.23610.20780.38040.2319
Avg. Item Coordination0.18400.35690.20710.20130.35240.22050.18780.33810.2061
w/o Cross-context Re-expression\color{red}{\text{w/o Cross-context Re-expression}}0.2011\color{red}{0.2011}0.4017\color{red}{0.4017}0.2289\color{red}{0.2289}0.2113\color{red}{0.2113}0.3712\color{red}{0.3712}0.2334\color{red}{0.2334}0.1926\color{red}{0.1926}0.3455\color{red}{0.3455}0.2117\color{red}{0.2117}
FedSG0.20580.40410.23540.21920.39440.24500.21860.39350.2433

Response to Q3: Evaluation protocol clarification

Yes. All LLM-generated user preference profiles and pseudo-reviews are constructed exclusively from training interactions and their associated reviews. Specifically, we first split each user's interactions chronologically into training, validation, and test sets using the leave-one-out protocol. We then generate user preference profiles and cross-context pseudo-reviews using only the training interactions and their associated reviews. Neither validation nor test interactions or their reviews are used in this process.

Reviewer_M4jqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes Federated Semantic Grounding (FedSG) for collaboration between independently maintained recommender systems without requiring known user or item correspondences. At the item level, shared semantic descriptors connect different catalogs, and their representations are trained on local interactions and selectively aggregated according to descriptor support. At the user level, an LLM re-expresses local preference profiles in another system’s item context, while preference mappings are coordinated across systems. The resulting representations are incorporated into local recommendation models. Experiments on three Amazon collaboration scenarios report improvements over independent and federated baselines, including settings with different backbones and collaboration between single-domain and cross-domain recommenders.

Critical strengths
  1. The work addresses collaboration between independent recommender systems and combines item semantics with user preferences to transfer knowledge without matched users or items.
Critical weaknesses
  1. Section 3.3 uses the number of items associated with a descriptor to guide semantic complementation. However, more items with few interactions may provide less useful behavioral evidence than fewer items with rich interactions. Eq. (6) triggers complementation whenever local support falls below a threshold, even if the other system has fewer items with that descriptor. It is unclear when such transfer helps. The threshold analysis does not resolve this concern, as it only shows how performance changes with the threshold, without establishing whether item counts reflect how well the descriptor representations have been learned.

  2. Eq. (11) requires gcrosspg_{\mathrm{cross}}^p to produce representations compatible with system pp, while glocqg_{\mathrm{loc}}^q serves system qq. Yet Eq. (9) directly combines these mappings, even though sharing a text input space does not ensure that their output spaces have the same semantic meaning. The compatibility of their dimensions is also unclear. Appendix D.3 specifies hidden dimensions of 64 for C2DSR and 128 for SyNCRec, but the paper does not specify the dimensions of the preference mappings or explain how they are made compatible across these backbones.

Questions for authors
  1. Cross-context re-expression preserves the user’s price sensitivity and quality expectations across domains. How does the method handle domain-dependent preferences, such as a user who prefers premium electronics but budget sports equipment?

  2. Appendix G.2 reports that Home performs best at ρp=0.7\rho^p=0.7, while Kindle favors a smaller value, although both belong to the same system. How was ρp\rho^p selected for the main experiments? Were both domains evaluated using the same value?

4 comments
AuthorsOpenReview ↗

Response to Reviewer M4jq (Part 1)

Response to Weaknesses 1: Clarification of Descriptor Support and Semantic Complementation

We thank the reviewer for the questions concerning how descriptor support is measured and how it guides semantic complementation. We clarify these two points below.

Regarding the first question, a descriptor represents shared semantic characteristics across different items. Learning from more associated items helps its embedding capture their common characteristics while reducing dependence on individual items. Therefore, the number of associated items provides a direct and interpretable measure of the breadth of item-level support for shared semantics.

The number of interactions can provide additional evidence about the amount of behavioral supervision. However, more interactions do not necessarily provide broader support for shared semantics, as they may be concentrated on a few popular items. Simply summing interactions could therefore overestimate the support from these items, while alternative weighting schemes require determining how much each item's interactions should contribute. We thus retain item count as a direct and interpretable measure, while recognizing the potential value of incorporating interaction counts.

Regarding the second question, in Eq. (6), the local number of associated items determines whether a descriptor needs complementation. The collaborating system's support is also considered in Eq. (5), where lower support leads to a smaller contribution to the aggregated embedding. Nevertheless, even with lower support, the collaborating system may still contribute useful knowledge from different items and interactions, making complementation potentially beneficial. The ablation results in Table 4 further show that selective complementation consistently outperforms uniform averaging across all three scenarios, supporting the effectiveness of this strategy.

Response to Weaknesses 2: Clarification of Semantic Coordination and Dimensional Compatibility

We thank the reviewer for raising this important question. We clarify the rationale for semantic coordination and the corresponding dimensional compatibility.

For the rationale for cross-system semantic coordination. Cross-context re-expression transforms user preferences from system pp into the item context of system qq. Thus, both gcrosspg_{cross}^p and glocqg_{loc}^q model preferences expressed in system qq's context. This connection motivates transferring the preference interpretation learned by system qq to guide cross-context preference modeling in system pp. Subsequent local optimization helps accommodate differences between independently learned representations, enabling the transferred knowledge to benefit system p's recommendation task. Furthermore, the "w/o Preference Coordination" ablation in Table 4, which removes cross-system preference mapping coordination, consistently underperforms FedSG across all three scenarios, further validating this design.

Regarding dimensional compatibility, all user profiles are first encoded into a shared 768-dimensional semantic space and then projected to h-dimensional representations. The 64 and 128 dimensions reported in Appendix D.3 correspond to separate homogeneous experiments using C2DSR and SyNCRec, respectively. For heterogeneous collaboration between C2DSR and SyNCRec, both systems use (h=64), ensuring identical parameter dimensions for the mappings coordinated in Eq. (9). We will clarify these settings in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer M4jq (Part 2)

Q1: Cross-context re-expression preserves the user’s price sensitivity and quality expectations across domains. How does the method handle domain-dependent preferences, such as a user who prefers premium electronics but budget sports equipment?

Response:

Cross-context re-expression aims to capture general user characteristics that can be transferred across autonomous recommender systems. We focus on broadly applicable evaluation criteria and decision-making tendencies, rather than introducing detailed preference rules for individual users or product categories. While the context-dependent preferences highlighted by the reviewer may exist, explicitly accounting for such differences would require additional assumptions and could introduce unreliable information into the generated profiles. To better handle users with different price and quality preferences across product categories, more targeted prompt design could be explored. The prompt in Appendix B.3 adopts a general approach, while refined instructions could explicitly distinguish such differences when supported by observed user behavior.

Q2: Appendix G.2 reports that Home performs best at (\rho^p=0.7), while Kindle favors a smaller value, although both belong to the same system. How was (\rho^p) selected for the main experiments? Were both domains evaluated using the same value?

Response:

We thank the reviewer for highlighting the need to clarify the distinction between parameter selection and sensitivity analysis.

In the main experiments, hyperparameters are selected at the system level, with Home and Kindle sharing the same ρp\rho^p. Specifically, we perform grid search while holding the collaborating system's parameters fixed, selecting the configuration that maximizes the average validation performance across the two domains.

The sensitivity analysis in Appendix G.2 provides a finer-grained, domain-level examination of this parameter. Specifically, we vary ρp\rho^p while fixing the other hyperparameters and separately examine its impact on Home and Kindle. The different preferred values reflect domain-specific responses to preference coordination, suggesting that more fine-grained, domain-adaptive coordination may offer further improvements.

Thank you for the clarifications. The rebuttal addressed some of my concerns, and I will maintain my positive assessment.

AuthorsOpenReview ↗

Thank You for Your Feedback and Support

Thank you for your feedback and continued support! We appreciate your suggestions and will take them into account when revising the manuscript.

Reviewer_yf6ZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Federated Semantic Grounding (FedSG) for collaboration between recommender systems without known user or item correspondences. On the item side, an LLM constructs shared semantic descriptors, and the systems selectively aggregate descriptor embeddings according to local catalog coverage. On the user side, FedSG re-expresses preference profiles in another system’s item context and coordinates the associated preference mappings. Experiments cover three Amazon collaboration scenarios, two main recommendation backbones, and heterogeneous configurations involving SASRec. The reported comparisons and ablations suggest that both coordination pathways improve recommendation performance in the tested settings.

Critical strengths

The controls and ablations provide substantive evidence that the proposed coordination adds value beyond semantic augmentation alone. The semantic-enhanced baselines use the same local semantic components (Appendix D.2), yet FedSG consistently improves the reported results across three scenarios and two backbones. For example, on Video with C2DSR, NDCG@10 increases from the strongest baseline’s 0.2739 to 0.3068 (Table 1). Removing either coordination pathway or replacing selective item complementation with averaging reduces performance across all three scenarios (Table 4). These comparisons are consequential because they strengthen attribution of the gains to the paper’s specific collaboration design.

Critical weaknesses

The semantic preprocessing cannot be adequately reproduced or audited. The paper does not identify the LLM or text encoder used to generate its semantic inputs. It also does not explicitly state whether held-out interactions were excluded when constructing user profiles from “interaction history.” Those profiles feed directly into recommendations, so this omission makes the reported gains harder to interpret.

Questions for authors
  1. Are all user profiles and generated pseudo-reviews constructed exclusively from training interactions and their reviews?
  2. Do the semantic-enhanced independent baselines retain the same cross-context inputs and model capacity as FedSG?
  3. Do the gains over the strongest matched baselines persist across random seeds and under full-catalog ranking or substantially larger candidate sets?
12 comments
AuthorsOpenReview ↗

Response to Reviewer yf6Z (Part 1)

Response to Weaknesses: Clarification of Semantic Preprocessing Details

We thank the reviewer for pointing out the need to clarify the semantic preprocessing details. We use Gemini-2.5-Pro as the LLM and all-mpnet-base-v2 as the text encoder, which produces 768-dimensional representations. The item-level and user-level preprocessing procedures are described below.

Item-level preprocessing. First, the LLM constructs a shared semantic schema consisting of high-level concepts and fine-grained descriptors based on the participating item domains. Next, the same LLM associates each item with predefined descriptors using its title, category, and brand. Finally, the text encoder converts these descriptors into embeddings, which initialize the learnable semantic representations.

User-level preprocessing. We select up to 15 recent training reviews per user. The LLM generates a user summary of approximately 50 words and six pseudo-reviews in the collaborating system's item context (three per domain). The text encoder then encodes the user summary together with the local reviews or generated pseudo-reviews to obtain local and cross-context user representations, respectively.

All semantic preprocessing is performed offline. The prompt templates are provided in Appendix B, and we will supplement the corresponding implementation details in the revised manuscript.

Q1 and Weaknesses: Are all user profiles and generated pseudo-reviews constructed exclusively from training interactions and their reviews?

Response:

Yes. All user profiles and generated pseudo-reviews are constructed exclusively from training interactions and their associated reviews. Specifically, we first split each user's interactions chronologically into training, validation, and test sets using the leave-one-out protocol. We then construct user profiles using only training interactions and their associated reviews, and generate cross-context pseudo-reviews accordingly. Neither validation nor test interactions or their reviews are used in this process.

Q2: Do the semantic-enhanced independent baselines retain the same cross-context inputs and model capacity as FedSG?

Response:

Yes. Sem-C2DSR and Sem-SyNCRec retain the same semantic inputs and model capacity as their corresponding FedSG implementations. Specifically, they use the same item semantic descriptors, local user preference profiles, and LLM-generated cross-context pseudo-reviews, processed by the same text encoder. Their semantic projection, preference mapping, and fusion modules also have identical architectures and parameter dimensions.

The only difference is cross-system semantic coordination: unlike these independent baselines, FedSG performs selective item-level semantic complementation and cross-system preference mapping coordination. Therefore, the comparisons isolate the benefits of semantic coordination rather than additional semantic inputs or greater model capacity.

Q3: Do the gains over the strongest matched baselines persist across random seeds and under full-catalog ranking or substantially larger candidate sets?

Response:

We appreciate the reviewer's suggestion for a more rigorous evaluation. In response, we substantially extended our experiments to cover all three collaboration scenarios and both recommendation backbones (C2DSR and SyNCRec), systematically examining performance stability across random seeds and under full-catalog ranking.

For random-seed robustness, we conducted three additional runs with different random seeds across all six scenario–backbone combinations. For full-catalog ranking, we further evaluated the models against the entire item catalog across the same six settings, using seed 42 to maintain consistency with the original experiments.

To ensure meaningful comparisons, we included the strongest matched baselines identified from our original results, including Sem-C2DSR/Sem-SyNCRec, Sem-FedAvg, Sem-FedAtt, Sem-FedMoE, and FFMSR, as applicable.

The additional results consistently confirm FedSG's advantages under both evaluation settings. Across all three new random seeds, FedSG consistently outperforms the strongest matched baselines across all collaboration scenarios, recommendation backbones, and evaluation metrics. Under full-catalog ranking, FedSG also maintains its leading performance across all settings. Although absolute ranking scores decrease as the candidate sets become substantially larger, FedSG's relative advantages remain consistent, indicating that the observed gains are not dependent on a particular random seed or the original sampled-ranking protocol. The detailed results are presented below.

AuthorsOpenReview ↗

Response to Reviewer yf6Z (Part 2)

1. Random-seed robustness (sampled ranking)\color{red}{\text{1. Random-seed robustness (sampled ranking)}}

Scenario 1\color{red}{\text{Scenario 1}}

Backbone: C2DSR\color{red}{\text{Backbone: C2DSR}}

SeedMethodVideo_MRRVideo_H10Video_N10Electronics_MRRElectronics_H10Electronics_N10Phones_MRRPhones_H10Phones_N10Sports_MRRSports_H10Sports_N10
43Sema-C2DSR0.26840.5273‾\underline{0.5273}0.3140.17380.35320.19750.2138‾\underline{0.2138}0.38410.22810.17790.34120.1991
Sema-FedAvg0.26440.52160.30990.1778‾\underline{0.1778}0.3582‾\underline{0.3582}0.2019‾\underline{0.2019}0.20470.37410.22750.18090.33750.2
Sema-FedAtt0.2734‾\underline{0.2734}0.5240.3174‾\underline{0.3174}0.17720.35790.20120.20540.3859‾\underline{0.3859}0.23140.17820.33250.1968
Sema-FedMoE0.25760.51750.3030.17270.33970.19260.20540.3859‾\underline{0.3859}0.2316‾\underline{0.2316}0.1833‾\underline{0.1833}0.3424‾\underline{0.3424}0.2027‾\underline{0.2027}
FFMSR0.24270.47110.28040.1610.31220.17790.19020.35330.21220.17620.32260.1933
FedSG0.28120.53450.32290.18530.36920.20490.21750.39400.23630.18720.34990.2047
44Sema-C2DSR0.262‾\underline{0.262}0.5149‾\underline{0.5149}0.3166‾\underline{0.3166}0.16720.33190.18730.20740.37950.23160.18340.33850.2052
Sema-FedAvg0.25320.49960.29490.1680.32490.18540.20830.37410.23060.18530.34060.204
Sema-FedAtt0.24590.49140.28670.16970.32490.18740.2123‾\underline{0.2123}0.3818‾\underline{0.3818}0.2371‾\underline{0.2371}0.1911‾\underline{0.1911}0.3427‾\underline{0.3427}0.2077‾\underline{0.2077}
Sema-FedMoE0.24260.50290.28780.17230.335‾\underline{0.335}0.1917‾\underline{0.1917}0.20430.3650.22520.18210.340.2016
FFMSR0.2510.48250.28940.1727‾\underline{0.1727}0.32960.19110.18880.34690.20940.17260.32380.1912
FedSG0.28190.53860.32410.18820.34570.20610.21420.39510.24390.19460.34690.2103
45Sema-C2DSR0.25830.51920.30430.1748‾\underline{0.1748}0.34080.19580.20440.38410.23030.17660.32850.1926
Sema-FedAvg0.25820.50370.29990.17240.34160.19350.21680.39030.23540.17730.34180.2011
Sema-FedAtt0.2621‾\underline{0.2621}0.524‾\underline{0.524}0.3084‾\underline{0.3084}0.17330.3447‾\underline{0.3447}0.1959‾\underline{0.1959}0.2202‾\underline{0.2202}0.3976‾\underline{0.3976}0.2365‾\underline{0.2365}0.18280.3474‾\underline{0.3474}0.2045
Sema-FedMoE0.2490.49140.28990.16770.32730.18670.19660.37140.22140.1879‾\underline{0.1879}0.34620.2081‾\underline{0.2081}
FFMSR0.24810.4760.28590.1610.3110.17720.18610.34240.20630.17190.31950.1889
FedSG0.27450.53630.32110.18170.35170.20250.22170.40220.23790.18970.35120.2109
AuthorsOpenReview ↗

Response to Reviewer yf6Z (Part 3)

1. Random-seed robustness (sampled ranking)\color{red}{\text{1. Random-seed robustness (sampled ranking)}}

Scenario 1\color{red}{\text{Scenario 1}}

Backbone: SyNCRec\color{red}{\text{Backbone: SyNCRec}}

SeedMethodVideo_MRRVideo_H10Video_N10Electronics_MRRElectronics_H10Electronics_N10Phones_MRRPhones_H10Phones_N10Sports_MRRSports_H10Sports_N10
43Sema-SyNCRec0.25310.57130.31190.16860.38260.19860.18470.39580.21590.1430.36110.1751
Sema-FedAvg0.25530.5762‾\underline{0.5762}0.31550.17360.39080.2050.18040.39040.21080.15110.36110.1816
Sema-FedAtt0.25910.56810.31620.16810.38730.19990.18760.40490.21990.1574‾\underline{0.1574}0.3766‾\underline{0.3766}0.1906‾\underline{0.1906}
Sema-FedMoE0.2622‾\underline{0.2622}0.56560.3178‾\underline{0.3178}0.177‾\underline{0.177}0.3978‾\underline{0.3978}0.2067‾\underline{0.2067}0.1976‾\underline{0.1976}0.4107‾\underline{0.4107}0.2296‾\underline{0.2296}0.15490.35490.1831
FFMSR0.24480.54280.29890.160.37680.19130.16050.37410.19080.13940.35050.1688
FedSG0.26690.58270.31870.18310.40310.20960.20740.42030.23380.16580.38470.1907
44Sema-SyNCRec0.27370.5754‾\underline{0.5754}0.32960.16370.3710.19230.19380.41970.2280.13580.3250.1589
Sema-FedAvg0.27460.56810.32870.16830.37490.1970.1943‾\underline{0.1943}0.4284‾\underline{0.4284}0.2344‾\underline{0.2344}0.1570.37350.1881
Sema-FedAtt0.2799‾\underline{0.2799}0.57290.3342‾\underline{0.3342}0.1680.3807‾\underline{0.3807}0.19820.18420.4130.22520.1638‾\underline{0.1638}0.3835‾\underline{0.3835}0.1962‾\underline{0.1962}
Sema-FedMoE0.27650.56070.32790.1701‾\underline{0.1701}0.37990.1999‾\underline{0.1999}0.18830.41660.22710.15040.340.1756
FFMSR0.25870.55750.31440.15880.36750.18780.17790.37860.20830.15020.32820.1745
FedSG0.28320.59490.33640.17660.38420.20830.20060.42930.23610.16700.38980.2017
45Sema-SyNCRec0.24740.5560.31240.16510.37920.19560.15440.38770.18940.15370.36540.1839
Sema-FedAvg0.2644‾\underline{0.2644}0.57130.3246‾\underline{0.3246}0.16680.38880.20.17550.38770.20640.17060.37170.2001
Sema-FedAtt0.25950.56970.31720.16260.38380.19530.17270.36590.1990.16860.36170.1961
Sema-FedMoE0.26210.577‾\underline{0.577}0.32110.1737‾\underline{0.1737}0.3993‾\underline{0.3993}0.2078‾\underline{0.2078}0.1918‾\underline{0.1918}0.4112‾\underline{0.4112}0.2339‾\underline{0.2339}0.1744‾\underline{0.1744}0.3804‾\underline{0.3804}0.2044‾\underline{0.2044}
FFMSR0.24930.53550.30030.16770.38340.19830.1830.40030.21860.14460.35550.1747
FedSG0.26920.57780.32580.17880.40540.21280.19350.41310.23410.17830.38480.2105

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviewer Discussion and Author Responses

We sincerely thank the Area Chair and all reviewers for their careful evaluations and constructive feedback. As the discussion period comes to a close, we would like to briefly summarize the key points of our discussions.

Our work investigates knowledge collaboration among autonomous recommender systems without requiring known user/item correspondences or homogeneous recommendation architectures. FedSG enables such collaboration through item-level semantic complementation and cross-context user preference coordination, while preserving the independence of participating recommendation models.

  • Reviewer yf6Z. The reviewer sought further details on semantic preprocessing and experimental settings, and asked whether the performance gains remain consistent across random seeds and under full-catalog ranking. We clarified the relevant implementation and evaluation details and conducted additional experiments with three new random seeds and full-catalog ranking, covering all three scenarios and both backbones. The results consistently support FedSG's performance advantages under these extended settings.

  • Reviewer M4jq. The reviewer raised questions about the rationale for item- and user-level semantic coordination, as well as hyperparameter selection and domain-dependent preferences. We clarified the design rationale and implementation of both coordination mechanisms, explained the hyperparameter selection protocol and sensitivity analysis, and discussed the treatment and limitations of domain-dependent preferences. The reviewer subsequently acknowledged that the rebuttal addressed some concerns and reaffirmed a positive assessment.

  • Reviewer NaPN. The reviewer asked how semantic components can be coordinated across independently trained, heterogeneous systems, and whether cross-context preference re-expression offers benefits beyond general semantic augmentation. We clarified the compatibility mechanisms and introduced a new controlled ablation to examine the specific contribution of cross-context re-expression. In the subsequent discussion, the reviewer acknowledged that our responses addressed their concerns to a considerable extent, while suggesting further clarifications in the revised manuscript.

We truly appreciate the thoughtful and constructive suggestions, which have helped us improve the clarity and completeness of our work. We will carefully incorporate these insights into the revised manuscript, and sincerely thank the Area Chair and all reviewers once again for their time and valuable contributions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.