acceptodds
Under review as a conference paper at ICLR 2027

Continuous-Time Sequential Recommendation via Personalized Gated Neural Flow

Abstract

Continuous-time sequential recommendation (CTSR) models continuously evolving user preferences from irregularly occurring interactions. Recent Neural ODE-based approaches show promise in capturing these dynamics, but their reliance on numerical integration significantly increase inference latency, entailing accuracy–efficiency trade-off. To address this limitation, we propose Personalized Gated Neural Flow (PGNF), a solver-free CTSR framework that directly parameterizes a flow map for continuous-time preference modeling. Through a temporal gating mechanism, PGNF evolves item states from a user's interaction history over time in a user-specific manner, then combines them to represent the user's evolving preferences. To complement these personalized dynamics with global collaborative signals, PGNF reflects information from an item co-occurrence graph into the initial item states. This graph captures relative temporal proximity between items within users’ interaction histories providing time-aware collaborative information without requiring additional computation cost at inference. Experiments on real-world datasets across diverse domains and scales demonstrate that PGNF achieves the state-of-the-art accuracy, improves computational efficiency, and substantially reduces inference latency compared to ODE-based approaches.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Sequential Recommendation, Neural Flows, Dynamical Systems

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

This paper studies continuous-time sequential recommendation and proposes Personalized Gated Neural Flow (PGNF). It first constructs a time-aware item co-occurrence graph to obtain initial item states xvx_v. For a user uu and target time t∗t^\ast, it then uses the history Sut∗S_u^{t^\ast} and the most recent interaction time suvt∗s_{uv}^{t^\ast} to compute a user-conditioned item state F(t∗,hv∣u)F(t^\ast,h_v\mid u) without numerical ODE integration. The evolved states are aggregated into a user representation eut∗e_u^{t^\ast}, and candidate items are ranked using ξ(u,w,t∗)=⟨eut∗,xw⟩\xi(u,w,t^\ast)=\langle e_u^{t^\ast},x_w\rangle. The paper establishes boundedness and invertibility onto the image under stated conditions and reports accuracy and efficiency results on three datasets.

Critical strengths
  1. The design addresses a practical computational limitation of continuous-time recommendation.Directly evaluating F(t∗,hv∣u)F(t^\ast,h_v\mid u) avoids numerical ODE integration, while the graph-enriched initial states xvx_v can be precomputed once training is complete. This gives a concrete mechanism for reducing inference cost while retaining collaborative information from A~\widetilde A. The batch-level measurements in Figure 2 support the efficiency of this design on Amazon-Tools across several state dimensions.

  2. The complete model achieves consistent improvements across datasets. PGNF obtains the highest reported scores across all metrics on Amazon-Tools, Behance, and Tmall-Buy. Its NDCG@20 scores are 7.88, 25.35, and 12.63, compared with 7.20, 24.17, and 11.56 for the strongest respective baselines. The chronological split, results averaged over five seeds, and standard deviations in the appendix strengthen this empirical evidence. Table 2 also provides useful evidence that the importance of collaborative information, temporal conditioning, and the readout differs across domains.

  3. The flow construction comes with explicit structural guarantees. Theorem 2.2 explains why the evolved state remains in Hd=(−1,1)d\mathcal H^d=(-1,1)^d, which is required for the subsequent tanh⁡−1\tanh^{-1} transformation. Corollary 3.1 extends the invertibility analysis to PGNF's two-branch candidate state and identifies concrete sufficient conditions on the gate scalers and operator norms. The distinction between invertibility onto the image and surjectivity over all of Hd\mathcal H^d, together with the restriction to intervals of fixed temporal context, makes the scope of these guarantees more precise.

Critical weaknesses
  1. The motivation and validation of individual design choices do not fully explain the method's complexity. Section 3 introduces graph initialization, temporal feature aggregation, two candidate-state branches, and alternative readouts, but the purpose of each addition and its connection to the overall design are not always clear. Table 2 includes a useful ResNet-Flow-style ablation, but does not isolate the contribution of the two branches in Eq. (17) or the invertibility constraints. Some temporal-component gains are also small: removing suvt∗s_{uv}^{t^\ast} changes Amazon-Tools NDCG@20 from 7.88 to 7.84, although the effect is larger on Behance. More targeted controls would help establish when the additional structure is necessary and what explains its benefits.

  2. The target-time information available to the compared models is insufficiently documented. PGNF explicitly conditions on t∗t^\ast through ϕ(t∗)\phi(t^\ast) and η(t∗,s,h)\eta(t^\ast,s,h). However, the experimental description does not specify which baselines receive this query-time information or how it is supplied. If t∗t^\ast comes from the held-out positive interaction, its availability in the intended serving scenario also needs explanation. These details matter for distinguishing architectural improvements from differences in temporal inputs and for interpreting the practical scope of the comparison.

  3. The connection between the theoretical assumptions and the trained implementation needs clarification. Corollary 3.1 requires strict bounds $|f_z|_L,|f_r|_L<1and $\|\widetilde\mu_{c_1}\|_L,\|\widetilde\mu_{c_2}\|_L<1/2. Section 3.3 states that spectral normalization and these constraints are used, but does not give the explicit scaling or parameterization that maintains the strict bounds during optimization. This matters because the selected α=2/5\alpha=2/5 and β=4/5\beta=4/5 satisfy α(16+5β)=8\alpha(16+5\beta)=8, placing the sufficient condition at its boundary. The conditional proof therefore does not, by itself, establish that the implemented model satisfies its assumptions.

  4. The efficiency study does not fully test the central accuracy-latency claim. Figure 2 reports batch-level measurements on Amazon-Tools while varying the state dimension, with the ODE baselines using a fixed three-step integration setting. This supports an efficiency advantage for that configuration, but does not show how the relative accuracy and runtime change with the solver's integration budget. A small sweep over integration settings would directly test the trade-off motivating the paper. The current measurements also leave scalability with respect to history length and item-catalog size less well established than scalability with respect to dd.

Questions for authors
  1. Target-time protocol. For a test instance (u,v,t∗)(u,v,t^\ast), is t∗t^\ast taken from the held-out positive interaction? Please clarify the intended serving scenario in which this query time is available. Which baselines receive t∗t^\ast or the corresponding elapsed-time features, and how are these inputs incorporated? A distinction between recommendation at a supplied time and prediction before the next interaction time is known would clarify the scope of the results.

  2. Theory and implementation. How are $|f_z|_L,|f_r|_L<1and $\|\widetilde\mu_{c_1}\|_L,\|\widetilde\mu_{c_2}\|_L<1/2 enforced during optimization? Please provide the explicit parameterization or scaling rule beyond the statement that spectral normalization is used. Since α=2/5\alpha=2/5 and β=4/5\beta=4/5 put α(16+5β)\alpha(16+5\beta) at the boundary value of 88, the strict operator bounds matter for applying the stated proof. I understand that the flow interpretation is restricted to intervals with fixed suvts_{uv}^t; this question concerns whether the trained model satisfies the assumptions within those intervals.

  3. Accuracy versus integration cost. Could the authors report accuracy and inference time for an ODE baseline under a small range of integration settings, using the same hardware and evaluation protocol as Figure 2? This would help assess whether PGNF improves the accuracy-latency trade-off beyond the reported three-step configuration.

6 comments

[W1] The motivation and validation of individual design components are insufficient. Some temporal components yield only small gains (e.g., removing suvts_{uv}^{t} changes NDCG@20 on Amazon-Tools only from 7.88 to 7.84).

Our ablations show that each component provides substantial benefits on at least one dataset, while retaining it does not appreciably degrade accuracy on the other evaluated datasets. These complementary contributions support PGNF as a shared architecture that works well across diverse recommendation settings without dataset-specific component selection.

Candidate-state branches. We added individual ablations of the two branches in Eq. (17) in Section 4.3 (Table 3). Removing either branch generally drops accuracy. The exception is c1c_1 on Behance, where performance remains comparable to the full model. This is consistent with Behance’s strong temporal variation and reliance on recency: temporal user context and gating have larger effects than ICG, suggesting limited additional benefit from c1c_1’s direct access to the initial item state. Conversely, the largest branch-ablation drop occurs when removing c1c_1 on Tmall-Buy (12.63 → 11.95).

Dataset-dependent contributions. We acknowledge that some gains are small, such as those from suvts_{uv}^{t} on Amazon-Tools (7.88 → 7.84 when removed). Given that Amazon-Tools spans approximately 15 years with short user sequences (8.08 interactions on average), we conjecture that an item’s latest interaction time provides limited information about the current intent. Behance and Tmall-Buy, on the other hand, span only about five months with denser interactions, and removing suvts_{uv}^{t} significantly drops NDCG@20 by 6.4% and 3.3%, respectively.

More broadly, removing ICG reduces NDCG@20 by 18.3% on Amazon-Tools; removing suvts_{uv}^{t} or TGM reduces it by 6.4% or 14.7% on Behance; and removing c1c_1 or c2c_2 reduces it by 5.4% or 3.7% on Tmall-Buy. These differences are consistent with the components addressing distinct temporal and collaborative characteristics across datasets.


ICG, TUC, and TGM denote the item co-occurrence graph, temporal user context, and temporal gating mechanism, respectively. Bold marks our default configuration.

ICGTUCTGM c1c_1TGM c2c_2Amazon-Tools NDCG@20Amazon-Tools HR@20Amazon-Tools MRRBehance NDCG@20Behance HR@20Behance MRRTmall-Buy NDCG@20Tmall-Buy HR@20Tmall-Buy MRR
✓✓✓✓7.88 ± 0.0416.49 ± 0.126.15 ± 0.0525.35 ± 0.0738.86 ± 0.2022.12 ± 0.0512.63 ± 0.1223.87 ± 0.1710.13 ± 0.15
✓✓✗✗7.62 ± 0.1715.85 ± 0.465.99 ± 0.0921.62 ± 0.0834.64 ± 0.2418.57 ± 0.1012.49 ± 0.1223.64 ± 0.2410.05 ± 0.09
✓✓✗✓7.80 ± 0.1016.35 ± 0.096.09 ± 0.1225.36 ± 0.1638.81 ± 0.2922.15 ± 0.1311.95 ± 0.3223.10 ± 0.269.49 ± 0.32
✓✓✓✗7.71 ± 0.0516.14 ± 0.156.04 ± 0.0325.17 ± 0.2238.65 ± 0.3821.96 ± 0.2012.16 ± 0.3823.38 ± 0.469.68 ± 0.36

[W2][Q1] It is unclear where the target time t⋆t^\star comes from, whether it is available in a realistic serving scenario, and which baselines receive query time or elapsed-time features and how.

Yes, as the reviewer guessed, query time is taken from the held-out positive interaction for each test instance (u,v,t⋆)(u,v,{t^{\star}}). PGNF constructs the user history using only interactions strictly before t⋆t^\star, excluding the target interaction itself and subsequent interactions. The positive item and negative candidates are scored at the same query time t⋆t^\star.

The intended serving scenario is recommendation when a user accesses the service or requests recommendations. The system observes the current timestamp when the request arrives, so t⋆t^\star is an available input feature at inference time. Elapsed times between past interactions and the request can also be computed immediately. In offline evaluation, we use the held-out interaction timestamps as query times to assess recommendation accuracy under this setting.

PGNF encodes the query time t⋆t^\star, the latest observation time ss of each historical item, and the elapsed time t⋆−st^\star-s. These temporal features are combined with the item state to compute the gates and candidate state of the personalized flow. The evolved item states are then aggregated through the readout to construct the user representation.

The baselines incorporate temporal information as follows:

BaselineTemporal inputs and their use
CTRecUses the elapsed time from the last interaction to t⋆t^\star for state decay and the intervals between past interactions and the query time for scoring.
TiSASRecUses pairwise time intervals within the observed history, without separately incorporating t⋆t^\star or the elapsed time to the query.
CTSMAIncludes t⋆t^\star when computing time intervals used by the intensity component of self-modulating attention.
TGSRecUses t⋆t^\star as the cutoff for temporal-neighbor retrieval and encodes elapsed times to neighboring interactions for attention.
GNG-ODEIntegrates the ODE up to the normalized query time t⋆t^\star to compute the historical interaction graph’s state.
GDERecUses temporal information in the training graph, without incorporating a query-specific t⋆t^\star during scoring.
TGODEIncorporates the normalized target time into ODE-based graph representation computation.
SS4RecUses interaction intervals, including the interval from the last interaction to t⋆t^\star, for time-aware SSM state updates.
SASRec, FEARec, BSARec, TV-Rec, KERecUse the item sequence observed before t⋆t^\star, without explicit query-time or elapsed-time features.

We follow each baseline’s originally proposed use of temporal information as closely as possible, preserving its intended design for a fair comparison.

Our method does not predict when the next interaction will occur. Instead, we evaluate recommendations given the time at which the user makes a request, which is the standard evaluation setting for continuous-time recommendation.

[W3][Q2] It is unclear how the strict Lipschitz bounds required by Corollary 3.1 are enforced during training beyond spectral normalization, which matters because α=2/5\alpha = 2/5 and β=4/5\beta = 4/5 yield α(16+5β)=8\alpha(16 + 5\beta) = 8, exactly at the boundary of the sufficient condition.

We added the implementation details and empirical verification to the Appendix B. In our implementation, we apply a margin factor κ=0.9\kappa=0.9 to spectral normalization. Since spectral norms are estimated approximately, the strict Lipschitz conditions may be violated during training. With this margin, however, the upper bound on the residual’s Lipschitz constant remained below 1 at every evaluated epoch.

Specifically, we normalize each layer’s weight as

\qquad \kappa_z=\kappa_r=\kappa,\quad \kappa_{c_1}=\kappa_{c_2}=\frac{\kappa}{2}.$$ Here, $W_*$ is the learnable weight, $\widetilde W_*$ is the weight used by the model, and $\widetilde\rho(W_*)$ estimates its spectral norm using one power iteration per training step. The target norms are therefore 0.9 for the update/reset layers and 0.45 for both candidate branches, below the required limits of 1 and 1/2, respectively. To verify this empirically, we compute exact layer norms using SVD after each epoch and substitute them into the corollary’s derivation to bound the residual’s Lipschitz constant. This bound remained below 1 across all datasets and evaluated epochs, with a maximum of 0.9595. Although the gate scaling factors $\alpha=2/5$ and $\beta=4/5$ yield $\alpha(16+5\beta)=8$, Corollary 3.1 still holds under the strict layer-wise conditions, which make the residual’s Lipschitz constant strictly smaller than 1.
Reviewer_GTyHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper argues that existing Neural ODE approaches for CTSR suffer from an accuracy–efficiency trade-off due to the mechanism of ODESolve. To address this, it proposes PGNF, a solver-free framework that directly parameterizes a flow map representing ODE solutions as a function of the initial item state and time, thereby avoiding repeated numerical integration. Its key idea is to model continuous-time user preference dynamics through personalized item dynamics: each item is evolved under a user-conditioned flow map, with the temporal user context injected into the gates so that the same item evolves differently across users. Initial item states are enriched with global collaborative signals via a time-aware item co-occurrence graph, and the evolved item states are combined through a readout layer into a time-specific user representation.

Critical strengths
  1. The paper is well-structured and easy to follow.
  2. The motivation of PGNF is clear and reasonable. The paper identifies the issue of the accuracy–efficiency trade-off in applying Neural ODEs to CTSR, and tries to solve it with the design of a parameterized flow map.
  3. The personalized item dynamics are theoretically grounded rather than purely heuristic, with several supporting theorems.
Critical weaknesses
  1. The analysis of the scalability is insufficient. Section 4.3 only reports the basic costs of baselines w.r.t. different state dimension sizes. If the authors can report the accuracy of baselines w.r.t. different state dimension sizes like Figure I, this experiment will be more persuasive in showing the scalability of PGNF.
  2. A case study is missing. The central claim of this paper is that user preference evolves in continuous time, yet all evaluations are aggregate ranking metrics, with no example showing that users with identical or similar interaction sets but different timestamps receive different recommendations. Therefore, a corresponding case study is needed.
  3. In the ablation study, the contribution of the direct branch c1c_1 and the reset-modulated branch c2c_2 is not studied.
Questions for authors

Could the authors conduct the experiments mentioned in the critical weaknesses? If the concerns are well addressed, the paper is worth getting a raised rating.

2 comments

[W1] The scalability analysis is insufficient, as Section 4.3 reports only the computational cost across state dimensions without the corresponding accuracy of the baselines.

Following your suggestion, we measured the NDCG@20 of PGNF and the baselines in Figure 2 at each state dimension dd. We added the results as a new rightmost panel of Figure 2 in Section 4.3. PGNF achieves the best or near-best NDCG@20 at every dd. We note that PGNF's efficiency does not come at the expense of accuracy, which makes our scalability analysis more robust. We sincerely appreciate your valuable advice, which has substantially improved our paper.


[W2] A case study is missing to show that users with identical or similar interaction sets but different timestamps receive different recommendations, which is central to the continuous-time claim.

Reflecting the suggestion, we added a case study in Section 4.6 (Figure 5) using Amazon-Tools. For the same interaction sequence, PGNF produces different recommendations as the target time changes. These changes are interpretable in that the recommendations reflecting long-term preferences remain relevant, while those associated with recent interests become less prominent as time elapses.

We further quantify this effect across three datasets in Table 4, copied below. We keep each interaction history fixed, including item identities, order, and historical timestamps, and vary only the recommendation time. Here, Δt\Delta t denotes the time since the latest interaction. We report the average percentage of shared items between the top-10 recommendations at each target time and those at the earliest target time for the corresponding dataset.

Amazon-Tools Δt\Delta tOverlap (%)Behance Δt\Delta tOverlap (%)Tmall-Buy Δt\Delta tOverlap (%)
1 day1001 min1001 day100
30 days9915 min692 days71
180 days9460 min4230 days56
365 days79180 min2290 days48

The decreasing overlap demonstrates that changing only the target timestamp changes the recommended items, even with an identical interaction history. Together, the qualitative case study and quantitative results support PGNF’s ability to capture time-dependent preferences.


[W3] The ablation study does not examine the individual contributions of the direct branch c1c_1 and the reset-modulated branch c2c_2.

We added individual ablations of c1c_1 and c2c_2 in Section 4.3 (Table 3), copied below. Removing c2c_2 drops accuracy across all datasets, supporting the value of reset-based modulation. Although both branches receive temporal context, c2c_2 additionally uses the reset gate to selectively weight the item-state features entering the candidate update.

Including c1c_1, which provides direct access to the initial item state, improves accuracy on the purchase-based datasets, while maintaining comparable performance on Behance. This is consistent with Behance’s recency-sensitive interactions: removing temporal user context or temporal gating causes a larger accuracy drop than removing the item co-occurrence graph, suggesting a stronger reliance on temporal adaptation.

Together, the two pathways allow PGNF to balance retention and adaptation across datasets with diverse characteristics. The additional direct branch improves accuracy where useful without an appreciable penalty on Behance. We further clarified this design rationale in Section 3.3.


ICG, TUC, and TGM denote the item co-occurrence graph, temporal user context, and temporal gating mechanism, respectively. Bold marks our default configuration.

ICGTUCTGM c1c_1TGM c2c_2Amazon-Tools NDCG@20Amazon-Tools HR@20Amazon-Tools MRRBehance NDCG@20Behance HR@20Behance MRRTmall-Buy NDCG@20Tmall-Buy HR@20Tmall-Buy MRR
✓✓✓✓7.88 ± 0.0416.49 ± 0.126.15 ± 0.0525.35 ± 0.0738.86 ± 0.2022.12 ± 0.0512.63 ± 0.1223.87 ± 0.1710.13 ± 0.15
✓✓✗✗7.62 ± 0.1715.85 ± 0.465.99 ± 0.0921.62 ± 0.0834.64 ± 0.2418.57 ± 0.1012.49 ± 0.1223.64 ± 0.2410.05 ± 0.09
✓✓✗✓7.80 ± 0.1016.35 ± 0.096.09 ± 0.1225.36 ± 0.1638.81 ± 0.2922.15 ± 0.1311.95 ± 0.3223.10 ± 0.269.49 ± 0.32
✓✓✓✗7.71 ± 0.0516.14 ± 0.156.04 ± 0.0325.17 ± 0.2238.65 ± 0.3821.96 ± 0.2012.16 ± 0.3823.38 ± 0.469.68 ± 0.36

Dear Reviewer GTyH,

Thank you again for your time and thoughtful feedback on our submission. Your comments have helped us strengthen the paper.

We especially appreciate your suggestion to include a case study. Conducting it was genuinely insightful for us: it clearly showed that our model produces different recommendations as time progresses, providing concrete evidence that it captures temporal dynamics as intended. We found this result both interesting and encouraging, and we believe it makes the paper considerably more convincing.

As the discussion period closes in about 10 hours, we wanted to check in briefly. We believe our rebuttal fully addresses every concern you raised, and the corresponding changes are in the revised manuscript.

If anything is still unclear, or you have further questions, please let us know. We are happy to clarify within the remaining time. If our responses have resolved your concerns, we would be grateful if you would consider updating your assessment.

Thank you once more for your constructive review.

Best regards, The Authors

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

Summary

This paper studies continuous-time sequential recommendation (CTSR), aiming to capture user preferences that evolve over time based on interactions that occur at irregular intervals. The goal is to predict which item a user may prefer at a particular future time, aligning better with the uneven temporal patterns. The authors introduce Personalized Gated Neural Flow (PGNF), whose central idea is to represent continuous-time preference dynamics in a personalized manner.

Critical strengths

Strengths

  • The paper is straightforward
  • The authors provide thorough mathematical detail, especially in Section 3.3.
  • Experiments are conducted on widely used baselines and benchmark datasets.
  • According to Table 1, it suggests good performance, though it would be helpful to report whether improvements are statistically significant.
  • The discussion on scalability (Section 4.3) is a valuable addition.
Critical weaknesses

Weaknesses

  • In my opinion, the contribution is marginal, largely building on existing ODE-based methods.
  • Evaluating ranking against only 1,000 sampled negative items is suboptimal; ideally, evaluation should consider the full item set.
  • The datasets used are relatively small, although they're well-known ones. Larger-scale datasets (e.g., Netflix, MovieLens20M) would strengthen the empirical evidence.
  • Several proofs/details in the appendices feel redundant, and they unnecessarily increase the paper length.
  • Tables III and IV in Appendix D should be discussed alongside Table 1 on page 7 rather than in isolation.
Questions for authors

Questions

  • Overall, I’m unsure how substantial the contribution is, given the reliance on ranking against only 1,000 sampled negatives and the lack of clarity on whether the reported gains are statistically significant.
5 comments

[W1][Q1] The contribution appears marginal, as PGNF largely builds on existing ODE-based methods.

We respectfully disagree that PGNF’s contribution is marginal. While PGNF builds on established continuous-time modeling techniques, its contributions address the personalization and efficiency requirements of recommendation as follows.

(1) We formulate personalized item dynamics through a directly parameterized, user-conditioned flow. This is for the first time in literature, to the best of our knowledge. The temporal relevance of the same item can differ across users. PGNF captures this difference by defining a flow over item states and conditioning their evolution on each user’s latest interaction with the corresponding item. This formulation produces user-specific trajectories from shared initial item states, enabling a single flow to represent personalized continuous-time preferences without separate user embeddings. Each item state can be evaluated directly at the target time, avoiding numerical integration and graph propagation during temporal evolution.

(2) To the best of our knowledge, we are the first to define temporal co-occurrence relative to each user's interaction timescale. The same temporal gap can carry different meanings for users with different activity frequencies. We address this by defining item-pair proximity through normalized ranks of temporal gaps within each user’s history, making temporal associations comparable across users before aggregating collaborative signals. These signals enrich the shared initial item states, complementing personalized dynamics without requiring graph propagation at inference. This design combines global temporal associations with user-specific evolution while keeping inference efficient.


[Q1] It is unclear whether the improvements reported in Table 1 are statistically significant.

Reflecting this comment, we now added the statistical significance test for all reported improvements. Every entry in Table 1 is the mean and standard deviation over five random seeds. For each column of Table 1, we compare PGNF with the strongest baseline in each column. As the variance across seeds differs between models, we use a two-sided Welch's t-test, which does not assume equal variances.

PGNF is significantly superior to the strongest baseline in every column of Table 1, with p < 0.01. The same holds for the new full-ranking evaluation in Table 2 on Amazon-Tools, Behance, and Tmall-Buy. We added this analysis to the "Statistical Significance" paragraph in Section 4.2. Please see Table IV in Appendix G for the gain, t-statistic, and p-value of each comparison.


[W2][Q1] Ranking evaluation against only 1,000 sampled negatives is suboptimal, and evaluation over the full item set is preferred.

Our original choice of 1,000 sampled negatives followed the protocol of TGSRec, which introduced the continuous-time evaluation setting we adopt. In the original version, we followed the same setting for the sake of fair comparison.

As the reviewer pointed out, however, we agree that evaluating over the full item set is ideal. Reflecting this comment, we additionally report full-ranking performance in Table 2 of the revised paper, concluding that our method remains effective under full ranking as well.


[W3] The datasets are relatively small, and experiments on larger-scale datasets would strengthen the empirical evidence.

Following the suggestion, we are running an experiment on Yelp [1], containing a large number of users (500k) and items (237k). We expect that it will finish before the deadline. We will update the result in Table 2 and Table IV in Appendix G once it is done. We are also running an experiment on the suggested MovieLens 20M. Despite promising, we would not be able to report its final result before the author response deadline. We will include the result in camera-ready.

DatasetUsersItemsInteractionsAvg. Length (± SD)DensityTotal TimeFiltering
Amazon-Tools16,63810,217134,4768.08 (± 5.86)0.079%5371 days5-core
Behance23,72429,794687,07028.96 (± 60.30)0.097%158 days5-core
Tmall-Buy125,55358,0582,064,29016.44 (± 8.43)0.028%152 days10-core
ML-20M138,49315,45119,964,833144.16 (± 228.17)0.933%7386 days10-core
Yelp500,871237,95715,698,98131.34 (± 56.75)0.013%6308 days10-core

[1] Zhao, Wayne Xin, et al. "Recbole 2.0: Towards a more up-to-date recommendation library." ACM international conference on information & knowledge management (CIKM). 2022.

[W4] Several proofs and details in the appendices are redundant and unnecessarily increase the paper length.

We removed redundant and less necessary experimental details from the main text. The appendix now keeps only what is needed to check the paper’s reproducibility and theoretical consistency.


[W5] Tables III and IV in Appendix D should be discussed alongside Table 1 in the main text rather than in isolation.

Following the suggestion, we revised the paper so that Tables III and IV, which report standard deviations, are now part of the main text and discussed there.

Update. We added full-ranking experiments on Yelp, with results reported in Table 2 and discussed in Section 4.2 of the revised paper. PGNF achieves the best accuracy on Yelp, further supporting its effectiveness on larger-scale data.

For your convenience, we copy Table 2 of the revised paper below. All values are on a percentage scale and are reported as mean ± std over 5 seeds.

ModelAmazon NDCGAmazon HRAmazon MRRBehance NDCGBehance HRBehance MRRTmall NDCGTmall HRTmall MRRYelp NDCGYelp HRYelp MRR
BSARec1.65 ± 0.033.80 ± 0.051.28 ± 0.028.94 ± 0.1015.63 ± 0.117.22 ± 0.101.86 ± 0.043.49 ± 0.051.55 ± 0.052.68 ± 0.025.92 ± 0.032.13 ± 0.02
TV-Rec2.01 ± 0.054.59 ± 0.211.52 ± 0.048.52 ± 0.0815.29 ± 0.106.81 ± 0.092.04 ± 0.043.53 ± 0.071.76 ± 0.063.40 ± 0.027.23 ± 0.042.71 ± 0.02
KERec1.87 ± 0.034.15 ± 0.041.46 ± 0.038.75 ± 0.1215.72 ± 0.226.97 ± 0.091.78 ± 0.033.31 ± 0.041.50 ± 0.032.93 ± 0.036.50 ± 0.042.31 ± 0.03
GNG-ODE1.74 ± 0.083.89 ± 0.261.38 ± 0.056.94 ± 0.1812.49 ± 0.215.57 ± 0.161.70 ± 0.053.32 ± 0.071.40 ± 0.052.81 ± 0.015.94 ± 0.032.25 ± 0.01
SS4Rec1.72 ± 0.153.95 ± 0.181.31 ± 0.159.69 ± 0.1216.86 ± 0.207.83 ± 0.092.16 ± 0.023.44 ± 0.041.94 ± 0.033.77 ± 0.068.02 ± 0.092.99 ± 0.06
PGNF2.24 ± 0.045.03 ± 0.121.72 ± 0.0310.77 ± 0.0818.56 ± 0.118.72 ± 0.082.42 ± 0.024.21 ± 0.032.08 ± 0.034.28 ± 0.029.47 ± 0.043.29 ± 0.01

We are also currently tuning all models on ML-20M and will report these results as soon as they are available.

Thank you again for your thoughtful review. Your comments have made the paper considerably more robust. Please let us know if you have any further questions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

A brief summary of our responses

Dear Reviewers jrFR, GTyH, and FhJs,

We sincerely thank you once again for the time and care you have devoted to reviewing our submission. Your thoughtful comments have been truly invaluable, and we believe they have substantially strengthened our paper.

We have posted point-by-point responses under each of your reviews and uploaded a revised manuscript. For your convenience, we briefly summarize the main updates below.

Reviewer jrFR

  • Contribution: We clarified the two main contributions of PGNF: (i) a directly parameterized, user-conditioned flow that produces personalized item trajectories and can be evaluated at any target time without numerical integration, and (ii) temporal co-occurrence defined relative to each user's own interaction timescale.
  • Statistical significance: Using Welch's t-test over five seeds, PGNF is significantly better than the strongest baseline in every column of Table 1, as well as in the full-ranking setting on Amazon-Tools, Behance, Tmall-Buy, Yelp (p < 0.01; Section 4.2 and Table IV in Appendix G).
  • Full-ranking evaluation: In addition to the sampled-negative protocol of TGSRec, we now report full-ranking results (Table 2).
  • Larger-scale data: We added full-ranking results on Yelp (about 501K users, 238K items, and 15.7M interactions), where PGNF achieves the best accuracy. Experiments on ML-20M are in progress, and we will share the results as soon as they become available.
  • Presentation: We removed redundant material and moved the standard-deviation tables (formerly Tables III and IV) into the main text, where they are now discussed alongside the main results.

Reviewer GTyH

  • Scalability with accuracy: We added the NDCG@20 of PGNF and the baselines at each state dimension dd (new rightmost panel of Figure 2). PGNF achieves the best or near-best accuracy at every dd while remaining efficient.
  • Case study: We added a case study on Amazon-Tools (Section 4.6, Figure 5) and a quantitative analysis on three datasets (Table 4). With an identical interaction history, changing only the target time changes the recommendations (e.g., on Behance, the top-10 overlap decreases from 100% to 22% as the elapsed time grows from 1 to 180 minutes).
  • Ablation of c1c_1 and c2c_2: We added individual ablations of the two branches (Table 3) and clarified their design rationale in Section 3.3.

Reviewer FhJs

  • Design components: We added individual ablations of the two branches in Eq. (17) (Table 3) and discussed the dataset-dependent contributions of each component. Each component provides a substantial gain on at least one dataset (e.g., removing ICG reduces NDCG@20 by 18.3% on Amazon-Tools, and removing TGM reduces it by 14.7% on Behance).
  • Target-time protocol: We clarified that t⋆t^\star is taken from the held-out interaction, corresponding to the time of a recommendation request in practice, and documented how each baseline incorporates temporal information.
  • Theory and implementation: We described the margin-scaled spectral normalization (κ=0.9\kappa = 0.9) in Appendix B and verified with exact SVD that the bound on the residual's Lipschitz constant remains below 1 (maximum 0.9595) across all datasets and evaluated epochs.
  • Accuracy-efficiency trade-off: We extended the scalability analysis to sequence length and catalog size, and added an accuracy-efficiency study against GNG-ODE and TGODE (Appendix H).

We fully understand that this is a very busy period, and we are truly grateful for your time. If you could kindly take a moment before the discussion period closes to look over our responses and the revised manuscript, we would greatly appreciate it. If any concerns remain, we would be more than happy to provide further clarification or additional results. If you feel that your concerns have been addressed, we would be sincerely grateful if you would kindly consider reflecting this in your evaluation.

Thank you very much once again for your valuable feedback and support.

Best regards, The Authors

Update. Our ML-20M experiments are now complete, and PGNF also achieves the best accuracy there under the full-ranking protocol, consistent with our Yelp results. We kindly ask the reviewer to take these results into account (see Table V in Appendix I).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.