acceptodds
Withdrawn from ICLR 2027

Dream Sparse, Connect Dense: A Lightweight Graph Prototype Dreaming Network for Cold-Start Sequential Recommendation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Sleep replay and dreaming in the human brain reactivate sparse daytime experiences and recombine them into structured neural trajectories. This process consolidates memories and facilitates future behavior anticipation. In sequential recommendation, data sparsity and cold-start issues are equally prevalent. Yet existing methods either rely on costly auxiliary information to compensate for sparse interactions, or employ sophisticated architectures that incur heavy computational overhead. Drawing an analogy to the human dreaming mechanism, we propose a Lightweight Graph Prototype Dreaming network (LightGPD) that translates the dreaming process into a structured data augmentation framework to enhance sequential recommenders in cold-start scenarios. Specifically, we first introduce a prototype learning module that discovers abstract interest prototypes from item embeddings via subspace alignment, providing the fundamental “memory traces” for subsequent augmentation. We then devise a dreaming module that augments sparse interactions through four sequential phases: interaction replay, abstraction, hypothesis generation and verification, and dynamic graph reconstruction. This module selectively reinforces prototype-aligned interactions and discovers new connections based on learned interest patterns, enriching the sparse training signal without requiring auxiliary modalities or user attributes. Extensive experiments on three real-world datasets demonstrate that LightGPD outperforms nine state-of-the-art baselines in both accuracy and efficiency. The code and datasets are available at https://anonymous.4open.science/r/LightGPD

Primary area:
learning on graphs and other geometries & topologies
Keywords:
sequential recommendation, cold-start recommendation, graph neural networks, data augmentation, prototype learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_trZbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LightGPD, a lightweight graph-based framework for cold-start sequential recommendation without auxiliary information. It uses a GCN backbone to obtain user and item embeddings and learns latent interest prototypes with soft item assignments. A prototype-guided “dreaming” module performs importance-aware interaction replay, maps item sequences to prototype-level interest states, predicts interest transitions, and reconstructs the interaction graph by selecting top-TT prototype-aligned user–item connections. User representations from the original and reconstructed graphs are fused with prototype-level interests for next-item prediction. Experiments on three datasets report improvements over sequential and cold-start recommendation baselines.

Critical strengths
  1. The paper presents a coherent and effective prototype-guided augmentation framework for sparse sequential recommendation. The method is technically well motivated, and the ablation results indicate that the main components contribute consistently to the final performance.
  2. The empirical evaluation is relatively comprehensive. The paper reports consistent gains across three datasets and further includes ablations, efficiency analysis, robustness experiments, and a dedicated evaluation on short-history users, which strengthens the empirical credibility of the proposed method.
Critical weaknesses
  1. The distinctive methodological contribution of the dreaming mechanism is not sufficiently established. The ablations support the usefulness of individual components, but the necessity and distinctiveness of their specific coupling remain unclear. In particular, the interest-transition prediction in Eqs. (11)–(12) mainly serves as an auxiliary objective, while graph reconstruction in Eqs. (13)–(16) is driven by aggregated prototype assignments and affinity-based Top-TT edge selection. It is therefore unclear what the full dreaming mechanism contributes beyond prototype-guided graph reconstruction with auxiliary transition regularization. The existing ablations and additional baselines do not fully resolve this through matched reconstruction controls or sufficiently close comparisons.
  2. The evidence supports a narrower cold-start setting than the motivation suggests. The introduction motivates newly emerging users or items with sparse or absent histories, while the dedicated cold-start evaluation considers users with fewer than five visible interactions in their evaluation prefixes. Although Appendix A explains how this is compatible with 5-core preprocessing, it remains unclear whether the evaluated users or items are truly unseen during training and whether graph construction is restricted to the visible history. Thus, the current evidence mainly supports short-history user recommendation rather than strict unseen-user or unseen-item cold-start.
Questions for authors

1.In Eqs. (11)–(16), do the predicted next-interest states directly affect graph reconstruction, or only indirectly through the auxiliary KL loss? What is the specific role of hypothesis generation in the dreaming mechanism?

2.Have the authors compared the proposed prototype-guided reconstruction with a prototype-free reconstruction based on backbone user--item similarities under the same Top-TT edge budget and propagation depth?

3.What exact cold-start setting is evaluated in Appendix F? Are the evaluated users and target items seen during training, and is graph construction restricted strictly to interactions available before each evaluation prefix?

4.How accurate are the newly added Top-TT edges? Do they recover held-out future interactions better than simple popularity- or backbone-similarity-based edge selection under the same edge budget?

5.In Eq. (5), ϕ(⋅)\phi(\cdot) is nonnegative but ψ(⋅)\psi(\cdot) is unconstrained. How are nonnegative prototype assignments guaranteed before they are treated as probability distributions in Eq. (12)?

Reviewer_qjhcOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

The paper proposes LightGPD for sequential recommendation with sparse user histories. It combines graph representations with shared interest prototypes, importance-based replay, and an auxiliary task that predicts the next prototype state. The proposed graph reconstruction step adds candidate user-item connections, followed by further graph propagation and a fused recommendation head. Experiments on Electronics, MovieLens-1M, and KuaiRec report improvements over several baselines, with additional ablations, efficiency comparisons, and cold-start and noise evaluations.

Critical strengths
  1. Sharing interest prototypes across users is a reasonable way to help users with limited histories without requiring side information. The connection between prototype-level prediction and graph augmentation is interesting, and could be useful if the proposed mechanism is supported by the implementation and experiments.

  2. The reported evaluation covers several relevant aspects of the problem. Besides the main comparisons on three datasets, the paper includes component ablations, graph and prototype baselines in the appendix, and targeted cold-start and noise tests. This is a useful experimental scope, although the issues below currently prevent me from drawing firm conclusions from the results.

Critical weaknesses
  1. The released implementation does not appear to implement the central graph reconstruction mechanism described in the paper. In LightGPD/GPD_Model.py, dynamic_graph_reconstruction returns top-T item indices and weights (lines 93-103), but its return value is discarded in forward (line 163). The subsequent memory_consolidation consists of linear layers and tanh, without propagation on the reconstructed graph. I also could not find the residual graph update in Eq. 18. Thus, in this implementation, T has no path to the prediction logits or training loss, which makes it difficult to reconcile the code with the T sensitivity results in Fig. 4 and the graph reconstruction ablation. I checked both the anonymous repository and the submitted supplementary archive; their model and training files match. This could still be a release-version issue, but it needs to be resolved before the gains can be attributed to the proposed method.

  2. The released evaluation pipeline leaves two important questions about the reported results.

First, GPD_Main.py evaluates test_batches during training and uses the resulting score for early stopping, checkpoint selection, and learning-rate scheduling (lines 230-241). The provided preprocessing creates train and test files, but I did not find a separate validation split or an independent final test evaluation. This differs from the validation-based selection described in the paper. If the released procedure was used for the main table, the test set influenced model selection. If a separate held-out evaluation was used, the corresponding split and script would resolve this concern.

Second, the cold-start subset in Appendix F is defined by fewer than five visible historical interactions. Counting the histories in the supplied test files gives 5,355 such examples in Electronics, but zero in MovieLens-1M and zero in KuaiRec; the minimum history length in both latter datasets is 15. Table 7 nevertheless reports cold-start results for both. This is not an objection to using 5-core data: shorter evaluation prefixes can certainly be constructed. Rather, the additional truncation or sampling procedure needed to obtain these subsets is not specified sufficiently to reproduce the evaluation. It also matters whether interactions hidden from the sequence are removed from the graph, since otherwise the user representation may still use that history.

These observations concern the released artifacts; I have not rerun training and do not assume that they establish how the reported tables were produced. However, they affect the central method and its evaluation, so I cannot support acceptance without clarification. I would reconsider if the authors can identify the exact experimental version and resolve these discrepancies.

Questions for authors
  1. Is the submitted code the version used for Tables 1-2 and Fig. 4? Please point to the implementation where the reconstructed edges enter graph propagation and where Eq. 18 is applied. If the wrong version was released, identifying the correct version would be the most useful first step.

  2. What train/validation/test split was actually used for the main results? Does test_batches in the released training script serve as validation data, and, if so, where is the separate final test evaluation? Please clarify whether the reported test scores were evaluated only after checkpoint and hyperparameter selection were complete.

  3. How were the MovieLens-1M and KuaiRec cold-start subsets in Table 7 constructed, and how many examples do they contain? If histories were truncated separately, were the graph and user representations restricted to the same visible interactions?

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

This paper presents a framework called LightGPD for cold-start sequential recommendation. The LightGPD framework has a so-called dreaming module and a prototype learning module. The experimental results show the effectiveness of the proposed LightGPD framework.

Critical strengths

S1. The proposed LightGPD sequential recommendation framework tends to address the cold-start problem without relying on auxiliary information.

S2. The experiments are conducted on three datasets.

Critical weaknesses

W1: The novelty of the dreaming module is not established against existing augmentation and replay work, and the comparison set cannot show it. It seems that each of the four phases reduces to a standard operation. In detail, interaction replay is weighted sampling, abstraction is a table lookup into the assignment matrix, hypothesis generation is a single linear layer followed by softmax, and dynamic graph reconstruction is top-T edge selection with a residual update. It is better to clearly present the novelty of the dreaming module.

W2. In the theoretical part, the effects of the number of prototypes αthe residual update factor γ should be carefully analyzed. The interpretability of these selection is insufficient if the current conclusion is just drawn based on the experimental parameter tuning.

W3: Many SOTA baselines are missing. Without a clear comparison with the SOTA baselines (e.g., [1][2][3][4][5]), it is difficult to thoroughly evaluate the performance of the proposed LightGPD framework.

W4. The authors are encouraged to consider the item cold-start or full cold-start scenarios. The current version of the paper only evaluates the performance for user cold-start (users with fewer than 5 visible historical interactions), and does not cover item cold-start or full cold-start scenarios (users with no visible interactions), which leads to insufficient experimental validation of the application boundaries of the proposed LightGPD framework.

[1] LLM-FSGNN: LLM-Guided Feature-Structure Augmentation Graph Neural Network for Cold-Start Recommendation. ACM Transactions on Information Systems, 2026

[2] Swarm-Optimized Graph-Based Recommendation System for Cold-Start User Adaptation. The 4th International Conference on Mobile Internet, Cloud Computing and Information Security, 2026

[3] GenRecEdit: Adapting Model Editing for Generative Recommendation with Cold-Start Items. SIGIR 2026

[4] Intent Oriented Contrastive Learning for Sequential Recommendation. AAAI 2025

Questions for authors

W1-W4

Comments on the paper

7 comments

My biggest concern was that the original checkpoints were selected on the test set. The validation-only reruns address that, and the new ablation shows what graph reconstruction contributes once the reconstructed edges actually enter propagation. I’ll raise my score to 3.

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

Many thanks to the authors for the detailed rebuttal and additional experiments. After reading the response to other reviewers, I will keep my score as it is.

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

Related papers

Discussion (0)

Sign in to comment.