acceptodds
Under review as a conference paper at ICLR 2027

SEMKOOPREC: SEMANTIC-CONTROLLED KOOPMAN DYNAMICS WITH CANDIDATE-PRESERVING RANKING CLOSURE

Abstract

Sequential recommendation often entangles three roles: interaction history represents current behavior, the latest event redirects it, and ranking translates a forecast into discrete choices. Pretrained semantics usually remain static item features. We introduce a state–control–decision principle in which event-dependent semantic control drives contextual evolution and a bounded decision interface regulates its ranking effect. SEMKOOPREC realizes this principle with a frozen behavioral anchor, persistent–transient Koopman dynamics, multi-horizon representation closure, and candidate-preserving decision closure. Our analysis follows semantic-event perturbations through persistent and exponentially decaying responses to catalog-wide score error, yielding margin-certified candidate stability or a finite anchor-relative intervention bound. Under the comparison protocol shared by every method in Table 1, SEMKOOPREC achieves the best result on all 12 HR/NDCG metrics across three Amazon benchmarks, with HR@10 gains up to 36.50%. A separate strict history-masked evaluation yields positive gains over the matched anchor across five datasets and alternative semantic representations. Matched ablations support event control, persistent–transient evolution, and bounded decision closure, with fewer than 73K additional trainable parameters.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Recommender Systems, Sequential Recommendation, Next-Item Recommendation, Semantic Recommendation, Candidate Ranking, Language-Enhanced Recommendation, Koopman Dynamics

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

This paper studies how item semantics can be used to model changes in user context and improve sequential recommendation. The proposed framework, SEMKOOPREC, separates historical behavioral state, latest-event control, and ranking adjustment. A frozen MPT backbone provides the behavioral state, while semantic and collaborative representations of the latest item supply a control input to a learned Koopman transition with persistent and contracting transient components. The model is trained to align multi-step predictions with future contextual representations. A subsequent ranking module protects an anchor candidate core, limits candidate admissions, and reorders the selected set.

The experiments report improvements over 19 baselines and variants on three Amazon datasets under a shared comparison protocol and gains over matched anchors on five datasets under a separate history-masked protocol. Additional experiments examine control inputs, an equal-parameter MLP alternative, ranking features, alternative semantic representations, and computational cost.

Critical strengths

S1. The paper provides an explicit connection between contextual dynamics and discrete ranking decisions. Theorems 1–2 trace event perturbations and representation errors through the persistent–transient transition to item scores, while Theorem 3 bounds candidate-set changes relative to the anchor. This analysis gives the structural guarantees of the proposed architecture and gives a principled basis for limiting how a learned forecast modifies an existing recommendation.

S2. The mechanism-focused evaluation in Section 4.4 makes a useful effort to separate the effects of the proposed components. The matched-anchor comparison holds the backbone and representations fixed , control ablations test the importance of event alignment and information sources , and an equal-parameter MLP provides an alternative to the structured transition. These comparisons make the design choices more assessable than aggregate ranking gains alone and provide meaningful support for the approach.

Critical weaknesses

W1. The practical role of bounded candidate intervention is not sufficiently examined. The introduction explicitly identifies bounded ranking intervention as requirement C3. However, the experiments do not sufficiently demonstrate how these constraints affect candidate selection in practice or what benefits they provide beyond the reported ranking gains. The reported HR@20 does not show how candidate membership changes relative to the anchor, since different candidate sets can yield the same hit rate. Reporting candidate-set overlap, realized admission counts, and cases where intervention gains or loses the target item would clarify how these constraints operate in practice.

W2. Some consequential parts of the decision layer are insufficiently specified. Appendix B.2 introduces a reliability gate κt\kappa_t, described as being computed from the contextual state and its forecast, but does not give its functional form or explain how it is trained or calibrated. Appendix B.4 lists five groups of ranking features without specifying the individual features. Table 10 also reports dataset-specific decision-layer settings without explaining their selections or search ranges. These make it difficult to reproduce the method and assess how much its performance depends on feature engineering and dataset-specific tuning.

Questions for authors

Q1. Could the authors report how the bounded intervention operates in practice, including candidate-set overlap with the anchor, realized admission counts, and how often candidate changes gain or lose the target item? How do the protected-core size and admission budget affect the trade-off between preserving the anchor and improving ranking?

**Q2. **Could the authors clarify the implementation and tuning details discussed in W2? These details would help assess reproducibility and the extent to which the reported gains depend on feature design and tuning.

1 comment

Dear AC, PC, and Reviewers,

We sincerely appreciate the time and effort devoted by the Area Chair, Program Chairs, and reviewers throughout the evaluation process. We are particularly grateful for the constructive feedback and valuable insights that have helped us further clarify and strengthen our work.

As the discussion period is approaching its deadline, we would like to kindly invite reviewers who have not yet had the opportunity to respond to take another look at our rebuttal. We have carefully addressed the concerns raised and would greatly value any further feedback or questions.

We genuinely hope to engage in a constructive and friendly exchange before the discussion period concludes, ensuring that any remaining concerns can be thoroughly clarified. We remain available and would be more than happy to provide additional explanations if needed.

Thank you once again for your time, thoughtful consideration, and continued engagement.

Best regards,

The Authors

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

SEMKOOPREC is a sequential recommendation model built on a "state-control-decision" principle that separates historical behavioral state, event-driven semantic control, and discrete ranking intervention. It leverages a frozen Markovian Pre-Trained Transformer (MPT) to define the baseline behavioral state, while utilizing persistent-transient Koopman dynamics to model how new semantic item interactions evolve the user's ongoing context. To translate this continuous evolution into a discrete catalog ranking, the model employs a bounded decision interface that protects a trusted set of anchor items, bounds the admission of new forecast-supported candidates, and localizes order correction.

Critical strengths
  1. First, the paper is well-written and logically fluent. The authors explicitly lay out the detailed mathematical formulations for their proposed framework, which provides a clear, structured, and easy-to-follow presentation of the methodology。

  2. Second, the paper features an exceptionally thorough evaluation. The authors validate SEMKOOPREC across multiple domains under both standard comparison and strict history-masked protocols, consistently achieving state-of-the-art performance on all 12 reported metrics. Furthermore, the extensive ablation studies and semantic-source transfer experiments convincingly isolate and confirm the individual effectiveness of the control, transition, and decision modules.

Critical weaknesses

The sequential modeling is not the reviewer's main research focus, and the reviewer cannot spot any major weakness in this paper. Instead, the reviewer raises some minor concerns as follows.

The experimental results in Table 1 show that SEMKOOPREC brings performance gains more than +20% on serval datasets evaluated with H@10 and N@10. For example, on the Toys dataset, SEMKOOPREC outperforms FreqRec-S by ~37%. It will be appreciated if the authors could provide more intuitive insights or concrete case studies to explain where exactly this massive improvement comes from.

Questions for authors

See the weaknesses.

2 comments

Dear AC, PC, and Reviewers,

We sincerely appreciate the time and effort devoted by the Area Chair, Program Chairs, and reviewers throughout the evaluation process. We are particularly grateful for the constructive feedback and valuable insights that have helped us further clarify and strengthen our work.

As the discussion period is approaching its deadline, we would like to kindly invite reviewers who have not yet had the opportunity to respond to take another look at our rebuttal. We have carefully addressed the concerns raised and would greatly value any further feedback or questions.

We genuinely hope to engage in a constructive and friendly exchange before the discussion period concludes, ensuring that any remaining concerns can be thoroughly clarified. We remain available and would be more than happy to provide additional explanations if needed.

Thank you once again for your time, thoughtful consideration, and continued engagement.

Best regards,

The Authors

Dear Reviewers,

Thank you very much for your time and effort in reviewing our paper.

As today is the final day of the rebuttal period, we would sincerely appreciate it if you could take a moment to review our responses and provide any further feedback.

We have carefully addressed your concerns and hope that our clarifications and revisions have satisfactorily resolved the issues raised. If you find that your concerns have been adequately addressed, we would be truly grateful if you could consider updating your evaluation and potentially increasing your score.

Your feedback means a great deal to us, and we sincerely hope to hear from you before the deadline.

Thank you again for your valuable time and consideration.

Best regards, The Authors

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

SemKoopRec realizes this principle via a frozen behavioral anchor, persistent–transient Koopman dynamics, multi-step representation closure, and candidate-preserving decision closure. It formalizes sequential recommendation through a state–control–decision principle, distinguishing historical behavior states, semantic control of current events, and bounded interventions in discrete ranking.

Critical strengths

S1: The paper is rich in content, providing extensive experiments and addressing a meaningful problem. "Continuous prediction at the representation layer" and "discrete ranking at the output layer" are two objects that cannot be directly measured by the same standard. Existing work hard-connects them using an unconstrained adapter, making it unclear what prediction brings and what prediction is allowed to alter. The proposed method adds structural constraints to both ends, ensuring that the combined effect can be theoretically proven and bounded.

S2: Through the setting of C1–C3 in the Introduction, the paper decomposes the problem into three components for separate verification. To address the issue of incommensurability between error metrics and ranking evaluation, it leverages a certified robustness approach to propose Theorem 2, which is an interesting solution.

Critical weaknesses

W1: The paper lays out three requirements (C1–C3). Such claims should be accompanied by corresponding justifications and verifications. In my understanding, treating "semantics as control" is the core motivation of this work to address controllability in recommendation. However, relying solely on the ablation study in Table 6 makes it difficult to verify the advantages of this mechanism for the entire architecture from more empirical or theoretical perspectives. There should ideally be a theorem or a concrete visualization/demonstration showing how different semantics reshape the context (e.g., how embeddings passed to downstream modules change, and how the upper bound is controlled) before presenting the impact on final metrics in Table 6. Additionally, since Koopman dynamics are naturally linear, corresponding visualizations could be performed.

W2: The premise of Theorem 2 relies on Equation 26, but there is little validation regarding the reasonableness of this premise. Furthermore, as the core module, the decision closure layer can be viewed as an adjustable plugin within this framework. To better validate this work, an experimental verification should be provided: applying the decision closure layer to an existing semantics-augmented recommender to compare its performance gain against the proposed method.

Questions for authors

Semantic Control: Beyond Table 6, how does semantic control theoretically bound downstream representations? Could you provide empirical or visual evidence demonstrating trajectory evolution under different semantics?

Theorem 2 & Plug-in Utility: What justifies the premise in Equation 26? Additionally, how does decision closure perform when integrated into existing semantic-augmented recommenders?

2 comments

Dear AC, PC, and Reviewers,

We sincerely appreciate the time and effort devoted by the Area Chair, Program Chairs, and reviewers throughout the evaluation process. We are particularly grateful for the constructive feedback and valuable insights that have helped us further clarify and strengthen our work.

As the discussion period is approaching its deadline, we would like to kindly invite reviewers who have not yet had the opportunity to respond to take another look at our rebuttal. We have carefully addressed the concerns raised and would greatly value any further feedback or questions.

We genuinely hope to engage in a constructive and friendly exchange before the discussion period concludes, ensuring that any remaining concerns can be thoroughly clarified. We remain available and would be more than happy to provide additional explanations if needed.

Thank you once again for your time, thoughtful consideration, and continued engagement.

Best regards,

The Authors

Dear Reviewers,

Thank you very much for your time and effort in reviewing our paper.

As today is the final day of the rebuttal period, we would sincerely appreciate it if you could take a moment to review our responses and provide any further feedback.

We have carefully addressed your concerns and hope that our clarifications and revisions have satisfactorily resolved the issues raised. If you find that your concerns have been adequately addressed, we would be truly grateful if you could consider updating your evaluation and potentially increasing your score.

Your feedback means a great deal to us, and we sincerely hope to hear from you before the deadline.

Thank you again for your valuable time and consideration.

Best regards, The Authors

Comments on the paper

2 comments

Dear Area Chair and Program Chairs,

Thank you very much for your time and efforts in coordinating the review process.

As the discussion period progresses, we would sincerely appreciate your assistance in encouraging the reviewers to engage with our rebuttal and provide further feedback. We have carefully addressed the reviewers' concerns and would welcome the opportunity to clarify any remaining questions through constructive discussion.

We would be particularly grateful if the reviewers could reconsider their initial assessments in light of our responses and, where appropriate, update their scores accordingly.

Thank you again for your consideration and support in facilitating a fair and constructive discussion process.

Best regards,

The Authors

Dear Area Chair and Program Chairs,

We sincerely appreciate your time and attention to SEMKOOPREC (Submission #13782). We first summarize our rebuttal efforts and the absence of reviewer follow-up, then highlight the contributions relevant to the final decision.

During rebuttal, we addressed substantive questions with new experiments. To verify semantic control beyond ablations, we measured latent displacement and persistent/transient responses under five controls on three datasets. Shuffled controls yielded larger shifts but worse ranking quality, while aligned semantic–collaborative control achieved the strongest NDCG@10. This shows why event alignment, not displacement magnitude alone, matters.

To examine our theoretical bounds, we compared observed latent and catalog-wide score errors with their bounds over horizons 1, 2, and 4. We also evaluated margin-based Top-K certification rates, ranging from 10.8% to 46.2% across the reported settings. We clarified that certification concerns dynamic-score Top-K stability under a margin condition, not final ranking.

To test whether Decision Closure is genuinely useful beyond our own architecture, we added it independently to frozen UniSRec and RecFormer backbones. Across six dataset–backbone settings, NDCG@10 improved by 9.93%–11.44% relative to the original models and by 4.13%–4.86% relative to plain reranking. This demonstrates benefits independent of Koopman dynamics.

We also measured Top-20 overlap, admissions, target gains/losses, and rank changes on five datasets. Decision Closure retained 94.3%–97.0% of anchor Top-20 candidates while admitting only 0.60–1.14 new candidates on average. A protected-core study quantified the preservation–accuracy trade-off. Six case studies traced how targets moved from anchor rankings through Koopman scores to final positions. We also specified the reliability gate, 19 ranking features, and hyperparameter search spaces. All additions are incorporated into the revised manuscript.

We also wish to emphasize how closely these additions correspond to the individual reviews. The requested trajectory evidence was supplied through control-specific shifts and persistent/transient decomposition. The requested test of the theorem's premise was answered with measured errors, bounds, and certification rates. The requested external validation of closure was answered by the two frozen-backbone plug-in experiments. The requested intervention diagnostics were answered with candidate membership and target-level statistics. The request to explain large gains was answered with concrete rank trajectories, while missing implementation details were made explicit. In short, each major criticism received a specific empirical or technical response; none was dismissed without analysis.

Despite these extensive and targeted responses, none of the reviewers has engaged in follow-up discussion about our rebuttal. Despite repeated invitations, we do not know which concerns remain. We respect the demands on reviewers, but worry that initial scores may not reflect this new evidence. We ask ACs and PCs to assess the complete rebuttal and revised manuscript.

Beyond the rebuttal, we believe the paper offers a distinct and substantive contribution. The state–control–decision formulation separates historical behavior, latest-event semantic control, and bounded ranking decisions. It addresses the disconnect between static semantic features, continuous context forecasts, and discrete ranking interventions.

Our method realizes this formulation through a frozen behavioral anchor, persistent–transient Koopman evolution, multi-horizon representation closure, and candidate-preserving ranking closure. Theoretical analysis traces semantic-event perturbations into retained and decaying contextual responses, then connects representation error to catalog-wide score error and candidate stability. The interface protects trusted candidates while allowing bounded admissions and local reranking.

Our experimental protocols are explicitly distinguished. Under the shared comparison setup, SEMKOOPREC ranks first on all 12 reported HR/NDCG measures across three Amazon benchmarks, with HR@10 gains reaching 36.50%. Under a distinct strict history-masked protocol, it improves over matched anchors on five datasets and with alternative semantic representations. Matched ablations, transition comparisons, and plug-in tests isolate contributing mechanisms. The model adds fewer than 73K trainable parameters.

We are not asking for special treatment or for unresolved limitations to be overlooked. We ask only that our manuscript be judged on its full technical merits, including the rigorous efforts undertaken to answer each reviewer concern. We hope the absence of reviewer discussion will not prevent this evidence from receiving fair and independent consideration.

Thank you again for your service and thoughtful assessment.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.