acceptodds
Withdrawn from ICLR 2027

Graph Process Neurons for Sequential Recommendation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Graph neural network (GNN)-based methods have achieved remarkable performance on sequential recommendation tasks. They compress collaborative relations and item transitions into compact node representations and propagate them to neighboring nodes, thereby learning user interests from collaborative signals. However, such representations are static summaries of historical interactions. User preference is inherently time-varying, and this compression overlooks how interest rises and fades between observed interactions. To address this limitation, we propose a Graph Process Neuron (GPrN) for sequential recommendation. It represents the user's history as a continuous preference process over real time and lets each candidate induce its own reading of it. Specifically, we first organize the visible interactions into a query-specific preference process along an observable temporal coordinate, which characterizes how user preference evolves over time. Then, each candidate induces branch-specific scalar reading kernels over complementary temporal views of this process, allowing candidate-dependent evidence to contribute at different scales. A factorized implementation over a finite cosine basis supports matrix-based full ranking. The final score augments structural and sequential preferences with a candidate-specific process residual controlled by a bounded query-adaptive gate. Extensive experiments on three real-world datasets show that GPrN outperforms nine state-of-the-art baselines on sequential recommendation tasks. The code and datasets are available at https://anonymous.4open.science/r/GPrN.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Sequential Recommendation, Graph Neural Networks, Process Neural Networks, Continuous-Time Preference Modeling, Recommender Systems

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

This paper proposes the Graph Process Neuron (GPrN) for sequential recommendation. Leveraging the irregular temporal information in users' interaction histories, GPrN constructs a continuous process using a Gaussian smoothing function. This process is then vectorized with a cosine basis, which is designed to produce a distinct, candidate-specific user representation.

Critical strengths

[S1] In a time-aware manner, modeling user histories as a continuous process offers a fresh perspective.
[S2] The proposed model achieves fast inference time, which supports its scalability.
[S3] The motivation behind the design choice is well justified from the view of sequential recommendation, and the mathematical formulation built upon it is sound.
[S4] The ablation studies and hyperparameter analyses are thorough.

Critical weaknesses

[W1] The role of the continuous process appears limited.

  • The expressiveness of the continuous process is governed by the number of cosine bases KK and the temporal smoothness σ\sigma of the Gaussian smoothing function (Eqs. 5 and 7). However, the model is highly robust to both hyperparameters; even with K=2K = 2 or σ=0.16\sigma = 0.16 and 0.320.32, there is no substantial performance degradation.
  • This suggests that the performance gains may stem not from capturing fine-grained temporal dynamics through the continuous process, but rather from its interaction location μ\mu or from the additional learnable components that follow it (Eqs. 8, 9, 11, and 12).

[W2] The experimental results raise concerns about the validity of the proposed model.

  • The datasets used in the experiments are overly limited in both domain diversity and scale.
  • There is a concern that the baselines may be underfitted due to insufficient tuning (see Q1–Q3).
  • The choice of baselines is questionable.
    • Some baselines are not standard for sequential recommendation: TGT is a multi-behavior model, and SURGE is a CTR prediction model.
    • The general sequential recommendation baselines, such as GRU4Rec, SASRec, and BERT4Rec, are relatively outdated, whereas recent graph-enhanced and time-aware sequential recommendation models, which are directly relevant to the proposed method, are not included.
Questions for authors

The following observations raise concerns about whether the baselines were appropriately selected and sufficiently tuned, and thus whether the reported improvements faithfully reflect the contribution of GPrN.

[Q1] The w/o Reader variant, which removes all components that constitute the core contribution of GPrN, already outperforms all baselines on Beauty and Sports (its results on ML-100K are not reported). This makes it somewhat difficult to attribute the reported gains to GPrN itself. One possible explanation is that the baselines may not have been fully tuned, or that the comparison may not include sufficiently strong relevant models. Could the authors comment on this result?

[Q2] Among the baselines, classic models such as GRU4Rec and SASRec achieve the best performance on ML-100K, and SASRec achieves the best performance on Sports. Since many subsequent models have been reported to outperform these classic models under comparable settings, this result suggests that the selected baselines may either be weaker than expected or may not have been sufficiently tuned. Could the authors explain this result, and comment on whether the conclusions would still hold against stronger recent models?

[Q3] Although the reported post-processing statistics and leave-two-out protocol are consistent with prior work, the overall performance scale differs noticeably from that reported in previous studies on the same datasets. Could the authors clarify the source of this discrepancy (e.g., full ranking vs. sampled negatives, metric computation, or preprocessing details), and describe how the hyperparameters of each baseline were tuned?

[Q4] Why is λp\lambda_p necessary? It appears that its effect could already be controlled by αmax\alpha_\mathrm{max} alone. This seems to be supported by the reported results, where the optimal performance is consistently determined by the product λp⋅αmax=0.6\lambda_p \cdot \alpha_\mathrm{max} = 0.6.

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

This paper proposes GPrN, a graph process neuron for sequential recommendation. It addresses the problem that GNN-based methods compress user history into a static vector and ignore how preference changes over time. To solve this, the authors model the timestamped history as continuous preference processes under three views and let each candidate read them with its own kernel, which is added to a LightGCN and GRU backbone as a residual score. Experiments on three datasets show that GPrN outperforms nine baselines.

Critical strengths
  • The paper models user history as a continuous preference process read by candidate-specific kernels, while keeping full ranking efficient.
  • GPrN achieves clear improvements over nine baselines on all three datasets.
Critical weaknesses
  • Most of the gain comes from the backbone rather than the process reader. GPrN w/o Reader already outperforms all baselines by a large margin (e.g., 8.83 vs. 7.63 Recall@10 on Beauty), while the reader adds only 0.83 on Beauty and 0.25 on Sports.
  • After expansion, the process reading becomes a weighted pooling of history embeddings with cosine features of time gaps, followed by a bilinear score. This is close to existing time-aware and target-aware attention.
  • Closely related continuous-time and recent sequential models are missing, such as TGSRec, TiCoSeRec, CL4SRec, and DuoRec, and some baselines are unusually weak (e.g., EA-GPS and FEARec are far below GRU4Rec).
Questions for authors

Please refer to weakness.

1 comment

Dear Reviewer,

Thank you very much for the time and effort you devoted to this paper. 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_NPRPOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper models user preferences as a continuous process that evolves over time, rather than as static representations. The proposed framework consists of a structural-sequential backbone and a process reader. The backbone processes the user's interaction history to construct a query context, which is then used to generate candidate-conditioned reading kernels for each candidate item. To capture different temporal characteristics of the interaction history, the framework employs three branches: coarse, local, and selective. The final recommendation score is computed by combining the backbone score and the process reader score through a weighted residual fusion mechanism.

Critical strengths

S1. While many sequential recommendation models focus on how to effectively encode and understand user interaction histories, the ultimate objective is to predict scores for candidate items. From this perspective, I believe the main strength of this work is its ability to generate candidate-conditioned reading kernels, allowing the same interaction history to be processed differently depending on the candidate item.

S2. Effectively capturing both long-term and short-term user preferences is important in sequential recommendation. I believe that incorporating coarse and local branches to capture these different temporal characteristics is a reasonable approach to making effective use of the available historical information.

Critical weaknesses

W1. The authors acknowledge the additional computational overhead introduced by the process reader and attempt to mitigate this issue through matrix-based full ranking. Nevertheless, I believe further experimental validation on datasets with large item catalogs is necessary to demonstrate the scalability of the proposed method.

W2. I have significant concerns about the overall reliability of the experimental results. FEARec has been shown to be one of the top-performing methods in numerous previous studies[1,2], often achieving substantially better performance than SASRec, BERT4Rec, and GRU4Rec. However, in Table 1 of this paper, FEARec exhibits considerably lower performance than these baselines. This unexpected result makes it difficult to fully trust the overall experimental findings.

[1] SIGMA: Selective Gated Mamba for Sequential Recommendation (AAAI'25)

[2] An Attentive Inductive Bias for Sequential Recommendation beyond the Self-Attention (AAAI'24)

Questions for authors

Q1. Regarding W1, could the authors include additional experiments on datasets with large item catalogs? Since the datasets used in this paper are relatively small, I would like to understand how the inference-time overhead of the proposed method changes as the number of candidate items increases.

Q2. Regarding Weakness 2, is there any particular reason why the relative performance of the baselines, especially FEARec, differs substantially from the trends commonly observed in previous studies?

Comments on the paper

7 comments

Thank you for the rebuttal and the additional experiments. However, I still have an unresolved concern regarding W1 and Q1.

I wonder why the authors chose to artificially expand the candidate set by resampling existing item representations rather than evaluating the model on an actual large-scale dataset, given that several publicly available recommendation datasets contain substantially larger item catalogs.

My original concern was about scalability on actual large-scale recommendation datasets, rather than solely inference latency with an artificially expanded candidate set. Since GPrN relies on graph construction and propagation, I believe that evaluating scalability on a real large-scale dataset would provide more meaningful evidence, as the computational cost may depend not only on the number of candidate items but also on the size and structure of the underlying graph.

Could the authors clarify the rationale behind this experimental choice?

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

Thank you for the detailed responses and the revised version.

Authors resolved my concerns on W1, Q1, and Q4. I also appreciate the full re-audit of all nine baselines, with the provenance table and both code versions released. In addition, I stand corrected on SURGE: it is a sequential recommendation model, and I mischaracterized it as a CTR model in my review. I apologize for the confusion.

But there are still remaining concerns. The response attributes the difference in performance scale to
(i) full ranking without sampled negatives
(ii) chronological leave-two-out
(iii) timestamp-preserving preprocessing
(iv) the history mask.

However, it is unclear to me which of these factors would differ from prior work in a way that affects the baselines. Full ranking without sampled negatives, chronological leave-two-out, and masking previously interacted items are effectively standard practice in sequential recommendation.

As for the timestamp-preserving preprocessing, the timestamps are retained only because GPrN itself consumes them; the chronological order of each sequence must be preserved in any case, and prior preprocessing already relies on timestamps for this ordering.

If the paper had deliberately adopted a different protocol for a well-motivated reason, for example a global timestamp-based split, which would arguably suit a model that relies on absolute time, different numbers would be understandable. Since the paper instead adopts a protocol shared by many prior studies, its results should be comparable to them. I believe a study should at least be comparable to prior work, and any deviation in the experimental setup should be convincingly justified. The discrepancies above remain unexplained.

In addition, the datasets used in the experiments are relatively small, and their domains are limited to Amazon and MovieLens, which makes it difficult to judge how well the findings generalize. I appreciate the authors' plan to add larger datasets in the revision; however, I can only evaluate the results that are currently presented in the paper.

Finally, most of GPrN's improvement already comes from its backbone (GPrN w/o Reader). Authors' response treats this backbone as part of GPrN, but it appears to be merely a simple modification of EA-GPS. I would therefore ask the authors to explicitly acknowledge this relationship in the paper and to show that the gain of the process reader is not specific to this particular backbone.

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

Related papers

Discussion (0)

Sign in to comment.