acceptodds
Withdrawn from ICLR 2027

Beyond Alignment: Toward Structural Compatibility Between Semantic and Collaborative Spaces

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Semantic information is widely used to complement collaborative signals in recommender systems, especially to address data sparsity, cold-start, and long-tail recommendation. Existing semantic–collaborative integration methods mainly focus on how to incorporate semantic information through enrichment, alignment, or disentanglement, while paying limited attention to what item–item relations are preserved after fusion. This question matters because the two views organize items differently: collaborative relations are derived from user behavior and shaped by item popularity, leaving tail items with few reliable collaborative neighbors, whereas semantic relations are derived from item content and can provide these items with meaningful neighbors. Consequently, fusion involves a structural trade-off, in which preserving more of one view's structure may weaken that of the other. To study this issue, we introduce structural compatibility, which characterizes how well the relational structures of both views are preserved after fusion, and quantify it through three measures: collaborative-neighborhood, semantic-neighborhood, and hierarchical semantic preservation. We further propose Structure-Preserving Fusion (SPF), which injects semantic information into the collaborative space through a controlled residual pathway and preserves semantic structure via hierarchy-aware regularization, thus avoiding global cross-view alignment. Experiments on three datasets show that SPF consistently improves recommendation performance, with larger gains for sparse interactions and tail items. Further analyses reveal that stronger semantic preservation can substantially weaken collaborative-neighborhood preservation without performance gains, indicating that effective fusion requires balancing the preservation of both structures. Overall, our findings establish structural compatibility as a complementary principle for analyzing and designing semantic–collaborative integration.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Recommender Systems, Semantic-Collaborative Integration, Structural Compatibility, Semantic IDs

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5Fi6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Structure-Preserving Fusion (SPF) for semantic–collaborative integration. It combines Controlled Semantic Residual Fusion with Hierarchy-Aware Regularization based on Semantic-ID prefix neighborhoods. Experiments on three datasets show improved recommendation performance, while structural analyses examine the trade-off between semantic and collaborative structure preservation.

Critical strengths
  1. The paper investigates a relevant problem: which semantic and collaborative relations are preserved after fusion, offering a useful perspective beyond representation alignment.
  2. Experiments on three datasets and five collaborative backbones on Amazon-Books show improved overall and tail recommendation performance under the evaluated settings.
  3. The regularization analysis provides useful evidence that stronger semantic structure preservation does not necessarily improve overall recommendation performance, while overall and tail recommendation performance can respond differently.
Critical weaknesses
  1. The methodological novelty is limited. SPF mainly combines residual fusion with a neighborhood-level InfoNCE objective based on Semantic-ID prefix neighborhoods. Its conceptual advance over the complementarity perspective of Wang et al. (2026) and the structure alignment in DaRec (Yang et al., 2025) is insufficiently established.
  2. The ablation study is insufficient to validate the core contribution. Comparing SPF with Controlled Semantic Residual Fusion (RF) demonstrates the benefit of Hierarchy-Aware Regularization, but does not distinguish the contribution of Semantic-ID hierarchies from supervision based on semantic neighborhoods.
  3. The comparison with representation alignment is not fully convincing. RLMRec-Con uses only item-side distillation in this paper, whereas the original framework includes both users and items. The alignment-weight analysis also shows better recommendation performance on Amazon-Books with a smaller alignment weight than the default, limiting conclusions about representation alignment.
  4. Collaborative Structure Preservation (CSP) measures neighborhood preservation relative to a separately trained collaborative filtering model. However, lower CSP accompanies better recommendation performance in the ablation study. The structural metrics therefore do not clearly establish which collaborative relations are useful and should be preserved.
Questions for authors

Please refer to Weaknesses.

Reviewer_C1CuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper looks at how to combine semantic information with collaborative signals without forcing the two embedding spaces to become fully aligned. The authors argue that each space contains useful but different neighborhood structures. They propose three metrics to measure how well these structures are preserved, and introduce SPF, a simple residual fusion method with a contrastive loss based on Semantic-ID hierarchy

Critical strengths

The paper asks an important and under-explored question. Instead of assuming that semantic and collaborative embeddings should be fully aligned, it points out that the two spaces may contain different but complementary neighborhood structures. This is a useful way to think about multimodal recommendation, and the proposed CSP/SNP/HC metrics provide a concrete starting point for studying that trade-off

The method is simple and practically useful. SPF can be added to several common CF backbones with little arch change

Critical weaknesses
  1. The evaluation is somewhat circular. The same RQ-VAE is used to build the training targets and evaluate SNP/HC. Therefore, improvements in these metrics are partly expected. Independent semantic labels or embeddings are needed

  2. The main trade-off is partly built into the metrics. If the semantic and collaborative neighborhoods are very different, preserving more of one will naturally preserve less of the other. This weakens the paper’s central claim unless the authors compare against a normalized baseline

Questions for authors
  1. is the same RQ-VAE used for SPF’s input, structural loss, and SNP/HC evaluation? If so, how do you rule out circular evaluation?
  2. why is SPF tuned per dataset while RLMRec-Con uses a fixed alignment weight?
  3. do the conclusions still hold with an independent semantic reference and after normalizing CSP/SNP for the original overlap between the two spaces?
Reviewer_xK2HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces structural compatibility to examine how semantic and collaborative relationships are preserved after representation fusion. It proposes SPF, combining residual semantic fusion with Semantic-ID hierarchical regularization. Experiments on three datasets demonstrate improvements in overall recommendation performance, particularly under sparse interactions.

Critical strengths
  1. The paper presents a meaningful perspective on semantic–collaborative integration by examining relational structures beyond conventional alignment.
  2. The structural diagnostics, ablation studies, and cross-backbone experiments provide useful empirical evidence. The observed non-monotonic relationship between semantic preservation and recommendation performance is particularly interesting.
Critical weaknesses

W1. Limited conceptual and theoretical justification. While structural compatibility is presented as a general design principle, its formalization primarily consists of neighborhood-preservation metrics. Related work has also examined structural differences and the limitations of global alignment. Moreover, the trade-off between CSP and SNP is partly constrained by the overlap of the original top-k neighborhoods. It remains unclear what additional theoretical insights or generalizable principles the proposed framework establishes beyond these descriptive observations. W2. Insufficient justification of Semantic-ID hierarchical regularization. The structural loss explicitly preserves semantic-prefix relationships, while collaborative-neighborhood preservation is not directly optimized. Table 3 shows that adding this loss increases SNP and HC but decreases CSP across all datasets. Although these results support a structural trade-off, they do not establish whether the improvements specifically arise from Semantic-ID hierarchy or from generic semantic regularization. Comparisons with continuous-semantic kNN regularization would help clarify the distinctive contribution of the proposed mechanism. W3. Unclear reliability of CSP evaluation. CSP relies on independently trained CF-only reference representations. In Table 4, RLMRec-Con with zero alignment weight still exhibits relatively low CSP (0.283 on Books and 0.401 on Games), despite recommendation performance close to the CF-only backbone. This raises concerns about whether CSP primarily reflects fusion-induced structural changes or differences between independently trained representations.

Questions for authors
  1. What advantages does SID-based hierarchical regularization provide over conventional semantic-neighborhood regularization?
  2. Are the CSP conclusions robust to independently trained CF references and within-model neighborhood comparisons?
  3. What additional insights does structural compatibility establish beyond the inherent trade-off between preserving different neighborhood structures?

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

Related papers

Discussion (0)

Sign in to comment.