- 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 . For a user and target time , it then uses the history and the most recent interaction time to compute a user-conditioned item state without numerical ODE integration. The evolved states are aggregated into a user representation , and candidate items are ranked using . The paper establishes boundedness and invertibility onto the image under stated conditions and reports accuracy and efficiency results on three datasets.
Critical strengths
-
The design addresses a practical computational limitation of continuous-time recommendation.Directly evaluating avoids numerical ODE integration, while the graph-enriched initial states can be precomputed once training is complete. This gives a concrete mechanism for reducing inference cost while retaining collaborative information from . The batch-level measurements in Figure 2 support the efficiency of this design on Amazon-Tools across several state dimensions.
-
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.
-
The flow construction comes with explicit structural guarantees. Theorem 2.2 explains why the evolved state remains in , which is required for the subsequent 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 , together with the restriction to intervals of fixed temporal context, makes the scope of these guarantees more precise.
Critical weaknesses
-
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 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.
-
The target-time information available to the compared models is insufficiently documented. PGNF explicitly conditions on through and . However, the experimental description does not specify which baselines receive this query-time information or how it is supplied. If 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.
-
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 and satisfy , placing the sufficient condition at its boundary. The conditional proof therefore does not, by itself, establish that the implemented model satisfies its assumptions.
-
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 .
Questions for authors
-
Target-time protocol. For a test instance , is taken from the held-out positive interaction? Please clarify the intended serving scenario in which this query time is available. Which baselines receive 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.
-
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 and put at the boundary value of , the strict operator bounds matter for applying the stated proof. I understand that the flow interpretation is restricted to intervals with fixed ; this question concerns whether the trained model satisfies the assumptions within those intervals.
-
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
[W2][Q1] It is unclear where the target time 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 . PGNF constructs the user history using only interactions strictly before , excluding the target interaction itself and subsequent interactions. The positive item and negative candidates are scored at the same query time .
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 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 , the latest observation time of each historical item, and the elapsed time . 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:
| Baseline | Temporal inputs and their use |
|---|---|
| CTRec | Uses the elapsed time from the last interaction to for state decay and the intervals between past interactions and the query time for scoring. |
| TiSASRec | Uses pairwise time intervals within the observed history, without separately incorporating or the elapsed time to the query. |
| CTSMA | Includes when computing time intervals used by the intensity component of self-modulating attention. |
| TGSRec | Uses as the cutoff for temporal-neighbor retrieval and encodes elapsed times to neighboring interactions for attention. |
| GNG-ODE | Integrates the ODE up to the normalized query time to compute the historical interaction graph’s state. |
| GDERec | Uses temporal information in the training graph, without incorporating a query-specific during scoring. |
| TGODE | Incorporates the normalized target time into ODE-based graph representation computation. |
| SS4Rec | Uses interaction intervals, including the interval from the last interaction to , for time-aware SSM state updates. |
| SASRec, FEARec, BSARec, TV-Rec, KERec | Use the item sequence observed before , 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 and yield , 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 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.
[W1] The motivation and validation of individual design components are insufficient. Some temporal components yield only small gains (e.g., removing suvt 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 c1 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 c1’s direct access to the initial item state. Conversely, the largest branch-ablation drop occurs when removing c1 on Tmall-Buy (12.63 → 11.95).
Dataset-dependent contributions. We acknowledge that some gains are small, such as those from suvt 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 suvt 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 suvt or TGM reduces it by 6.4% or 14.7% on Behance; and removing c1 or c2 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.