acceptodds
Withdrawn from ICLR 2027

Same Items, Different Roles: Context-Aware Facet Disentanglement Graph Network for Discriminative Bundle Recommendation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Discriminative bundle recommendation aims to rank predefined item collections (i.e., bundles) for users. Multi-view graph methods integrate user-item, user-bundle, and bundle-item relations. However, they do not use co-member-conditioned facet roles and explicit bundle structure to dynamically select from history-derived interests when ranking candidate bundles. To address this gap, we propose a context-aware facet disentanglement graph network for discriminative bundle recommendation, namely CFDGN. First, a co-member-conditioned router combines intrinsic item facets with bundle context to infer bundle-specific roles. Second, a role-aware component summarizes role-derived bundle structure as three unlabeled soft states (i.e., Similar, Complementary, and Noise). Third, historical bundles provide candidate-independent interest values for each user. The current candidate selects their mixture, which is then evaluated through four structural matching paths. A positive pointwise mutual information (PPMI) prior derived exclusively from training-observed bundles supplies auxiliary relation evidence. A calibrated structural residual augments the backbone MultiCBR's collaborative score. Experiments on three benchmark datasets show that CFDGN exceeds eleven competing baselines on both Recall and NDCG. Its relative Recall@20 gains over MultiCBR are 3.8%, 1.7%, and 8.3%, respectively.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
bundle recommendation, recommender systems, graph neural networks, disentangled representation learning, multi-view learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BaQXOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CFDGN, a context-aware facet disentanglement graph network for bundle recommendation. It models bundle-specific item roles, structural bundle semantics, and candidate-conditioned user interests. Experiments on three datasets demonstrate consistent improvements over existing recommendation baselines.

Critical strengths
  1. The context-aware facet modeling provides an interesting approach to capturing bundle-specific item roles.
  2. Candidate-conditioned interest selection effectively incorporates user history and bundle structure.
  3. Experiments on three public datasets demonstrate consistent improvements over eleven baselines.
  4. The paper provides comprehensive architectural descriptions and useful mechanism-level analyses.
Critical weaknesses
  1. The semantic interpretations of the learned item roles and bundle states lack external validation.
  2. Component ablations are limited to a single dataset and seed, with relatively small performance differences.
  3. The PPMI component shows negligible improvement in the reported ablation study.
  4. The computational overhead of the additional structural modeling components is insufficiently evaluated.
Questions for authors
  1. Can the learned roles and bundle states be validated against human-annotated semantics?
  2. Can multi-seed ablation experiments be provided across all three datasets?
  3. Why does removing PPMI have almost no impact on recommendation performance?
  4. What are the training and inference costs compared with MultiCBR?
1 comment

Dear Reviewer,

Thank you very much for the time and effort you devoted to this paper during the discussion period. We have decided to withdraw the paper from OpenReview in order to protect the intellectual property involved. We sincerely appreciate all the valuable suggestions you raised. We wish you all the best.

Best regards,

The Authors

Reviewer_hGF4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the ranking of predefined item bundles and proposes CFDGN, which adds a structural residual branch to a MultiCBR backbone. The method infers contextual item roles from differences between an item's facet salience and that of its co-members, constructs bundle representations, and uses the candidate bundle to select among interests derived from the user's history. The authors report improvements over the selected baselines on three datasets.

The main concerns are that the released implementation uses inconsistent checkpoint-selection protocols, compromising the validity of the performance comparison, and that the experiments do not establish whether the gains come from the proposed contextual-role mechanisms rather than the additional representation and training capacity of the structural branch.

Critical strengths
  1. The complete model outperforms MultiCBR on the reported metrics across three datasets. The results on iFashion, in particular, merit further investigation.

  2. The information sources and scoring procedure of the structural branch are reasonably well specified. The released code makes it possible to inspect the correspondence between the method and the experimental protocol.

Critical weaknesses
  1. The released implementation does not support a fair, consistent comparison underlying the paper's central performance claims. The reported gains therefore cannot be attributed directly to the proposed method.

Appendix A.1 states that checkpoints are selected by maximizing the sum of validation Recall@20, NDCG@20, Recall@40, and NDCG@40. However, the released code applies a different rule to MultiCBR: it updates the checkpoint only when both Recall@20 and NDCG@20 exceed those of the previously selected checkpoint. CFDGN instead uses the four-metric sum.

These rules are not equivalent. The MultiCBR rule can reject a checkpoint with better aggregate performance if one metric decreases slightly, whereas the CFDGN rule can accept it. The released comparison thus changes both the model architecture and the selection procedure, and does not provide a controlled estimate of the benefit of the structural branch.

This issue bears directly on the main conclusion. On NetEase, the absolute improvements are only 0.0016 in Recall@20 and 0.0009 in NDCG@20. Without ruling out the effect of checkpoint selection, the evidence is insufficient to attribute these differences to the proposed mechanisms.

This is not an incidental implementation detail that can be resolved through a brief clarification alone. The authors need to identify the actual protocol used to produce the reported results and provide comparisons under a common selection rule. Until then, the central evidence for the method's claimed advantage remains inadequate.

  1. The paper does not establish the substantive value of its central innovations. The existing ablations instead show that most of the gain survives the removal of either core mechanism individually.

In Table 3, the full model improves Recall@20 over the backbone by 0.0111. Removing contextual-role routing reduces Recall@20 by only 0.0010, and replacing candidate-conditioned interest selection with uniform weights reduces it by only 0.0006. Thus, most of the overall gain remains after either of the two emphasized mechanisms is removed separately.

These results do not prove that the components are ineffective. They do, however, expose the central attribution gap: the experiments primarily establish that adding a structural scoring branch helps, not that the proposed role inference and interest selection are responsible for its effectiveness.

MeanPoolBundle does not resolve this issue because it replaces only one representation path while retaining the other structural features and matching mechanisms. There is no conventional pooling or attention-based residual branch with the same inputs, comparable parameter capacity, and the same training procedure. Additional capacity, use of historical information, and residual training therefore remain plausible explanations for most of the gain.

This gap also weakens the novelty claim. BRUCE already models items in the context of their bundle, HIDGN explicitly considers how the same item can express different intents across bundles, and MIND uses target-aware attention to select among historical interests. A different salience-contrast computation does not, by itself, establish a capability absent from these methods.

References:

Internal role-switching statistics cannot substitute for functional validation. The router normalizes the difference between item and context salience again, so a small difference can produce a substantial change in role probabilities. Observing a role switch therefore does not establish that the model has identified a useful contextual role, much less that such changes explain the ranking gains.

The paper currently establishes a particular implementation and its aggregate performance, rather than the necessity or advantage of its central modeling idea. Without that evidence, assembling these components does not constitute a convincing technical contribution.

Questions for authors
  1. Do the reported advantages persist under a common checkpoint-selection protocol?

Please specify the checkpoint-selection rules actually used for Tables 2 and 3, and compare MultiCBR and CFDGN under the same rule. Since the released implementation uses different rules, this requires clarification at the level of the results, not merely a code or wording change.

  1. On iFashion, where the overall gain is largest, do the proposed role mechanisms outperform a conventional additional branch?

Please include a standard pooling or attention-based structural branch with the same inputs, comparable parameter capacity, and the same residual training and calibration procedure. Also compare the four combinations of contextual-role routing enabled/disabled and candidate-conditioned interest selection enabled/disabled.

These comparisons should answer a specific question: can nearly the same improvement be obtained without the proposed role and selection mechanisms? If so, the paper's characterization of its central contribution requires substantive revision.

  1. What verifiable capability does this method add over directly related contextualization or intent-disentanglement methods?

Please prioritize a fair comparison with directly related methods such as HIDGN or BRUCE, accounting explicitly for differences in tasks and available supervision. The authors need to demonstrate a practical advantage of the specific salience-contrast mechanism. A different formula, or the absence of exactly the same combination of components in prior work, is not sufficient to establish a technical contribution.

  1. How can useful role changes be distinguished from sensitivity introduced by normalization?

Please evaluate, using an actual trained checkpoint, how small perturbations to contextual salience affect role distributions and final rankings, and compare variants with and without the second normalization of the difference vector. The relevant evidence is that role changes reliably improve the associated predictions, not merely that the dominant internal dimension can switch.

1 comment

Dear Reviewer,

Thank you very much for the time and effort you devoted to this paper during the discussion period. We have decided to withdraw the paper from OpenReview in order to protect the intellectual property involved. We sincerely appreciate all the valuable suggestions you raised. We wish you all the best.

Best regards,

The Authors

Reviewer_tSMYOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes CFDGN, a context-aware facet disentanglement graph network for discriminative bundle recommendation, addressing the limitations of existing multi-view graph methods in modeling bundle-specific item roles and dynamically selecting user interests. The framework integrates co-member-conditioned facet routing, role-aware bundle structure modeling, and candidate-conditioned interest selection to enhance structural matching between users and bundles.

Critical strengths
  1. The problem of bundle recommendation is valuable and important, regarding both academic research and industrial applications.
  2. Extensive experiments have been conducted.
Critical weaknesses
  1. The presentation could be substantially improved. The logics in the introduction should be clearer; the jargons need to be clarified; the table presentation should be better formatted, e.g., cell alignment, margins beside tables; the reference list should be formatter, e.g., some authors use full name, some use short-form abbreviation, the same problem applies to the conference name, etc.
  2. The performance improvement of the propsed method compared with SOTA is marginal.
  3. The paper is hard to follow, the motivation is more intuition-driven, without profound support; it is not clear how and why the proposed method can fully address the identified problems.
Questions for authors

please see the weakness above.

1 comment

Dear Reviewer,

Thank you very much for the time and effort you devoted to this paper during the discussion period. We have decided to withdraw the paper from OpenReview in order to protect the intellectual property involved. We sincerely appreciate all the valuable suggestions you raised. We wish you all the best.

Best regards,

The Authors

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

Related papers

Discussion (0)

Sign in to comment.