- 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?
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.
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.