- Rating
- 1
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper models timestamped user–item interactions as a counting process, decomposing item choice logits into time-varying popularity and user-specific utility. Theoretically, it proves that unmodeled popularity induces a shared directional shift across users that collapses representation uniformity and that zero-mean centering across users restores identifiability. Methodologically, it jointly trains a Hawkes popularity process and any recommender backbone via a 15-negative importance-sampled softmax loss plus a U-statistic centering penalty. While effective on synthetic and head items, unregularized Hawkes rates over-concentrate on KuaiRand dataset.
Critical strengths
By modeling timestamped implicit feedback via conditional intensity and conditioning on user interaction arrival times (Proposition 2), the paper cleanly eliminates unknown user activity rates and reduces temporal popularity confounding to an additive logit ambiguity. Lemma 1, Theorem 1, and Corollary 1 then provide a transparent geometric proof of why unmodeled temporal popularity causes representation collapse: projecting onto item embeddings induces a shared shift across all user-context vectors at time 𝑡, contracting pairwise distances on the unit sphere and driving the hyperspherical uniformity loss to 0. This theoretical mechanism is directly verified in the controlled 30-seed synthetic experiment
Rather than hiding or glossing over the sharp drop in Tail Recall@10 on KuaiRand in Table 1, the authors dedicate Appendix A.7 to a rigorous root-cause investigation. By comparing utility-only inference against default 𝜂=0.5 (Table 4), measuring normalized entropy 𝐻_norm and top-5% probability mass of the fitted Hawkes distribution vs. empirical 3-day popularity (Table 5), and plotting learned background rates across item frequency percentiles, the authors clearly separate the effectiveness of the learned utility from the over-concentration of the unregularized Hawkes background rate.
Critical weaknesses
Table 1 Reporting Errors:Duplicated SASRec/MF Row: In Table 1 on Micro-Video, the entire row for Vanilla SASRec (0.0121±0.0009, 0.0061±0.0006, 0.0168±0.0017, 0.0089±0.0009, 0.0112±0.0005, 0.0054±0.0005) is identical to Vanilla MF. Might be some copy/paste problem
Missing Loss-Function Control Standard sequential recommenders (SASRec, TiSASRec, GRU, etc.) are typically trained with binary cross-entropy (BCE) or BPR using 1 negative sample, whereas +(Ours) trains the backbone with an importance-corrected sampled softmax choice loss using M=15 negative samples and adds a separate 128-dimensional item embedding table. In collaborative filtering and sequential recommendation, replacing 1-negative BCE/BPR with multi-negative sampled softmax is well known to produce large gains in both Recall/NDCG and representation uniformity (since softmax cross-entropy directly optimizes an alignment-and-uniformity objective). Without an ablation training the vanilla backbone under ℓ_choice with uniform popularity (𝜆_𝑣 ≡ 1, 𝛾=0) it is impossible to tell how much of the large jump in Table 1, Table 2, and Figure 2 comes from temporal popularity modeling vs. simply switching the training loss to 15-negative sampled softmax.
Disconnect Between Theory and Practice on the Centering Penalty (𝛾) and Unregularized Hawkes Estimation (λ_v)
Negligible Empirical Impact of Centering: While Proposition 1 relies on user-centering to establish identifiability, Figure 2 shows that setting γ=0 (no centering penalty) already achieves essentially the same representation uniformity and Recall@10 as the best γ>0 across all six backbones and three datasets. This suggests that architectural separation alone already prevents most popularity absorption in practice.
Unregularized Joint Estimation of 𝜆_𝑣(𝑡;Φ): Inside ℓ_choice because 𝜆_𝑣(𝑡;Φ) is trained solely through the 15-negative importance-sampled softmax choice loss ℓ_choice rather than a temporal point-process likelihood, the background rate is poorly calibrated on KuaiRand, concentrating 99.4% of probability mass onto the top 5% of items (Table 5). When combined at inference via Eq. (3) using the validation-selected 𝜂=0.5 (chosen on overall Recall@10, which favors Head items), it destroys Tail Recall@10 by 85%–87% across GRU, SASRec, TiSASRec, and FEARec.
Questions for authors
see above weakness summary
8 comments
Table 2-1: Micro Video
| Method | All Recall@10 | All NDCG@10 | Head Recall@10 | Head NDCG@10 | Tail Recall@10 | Tail NDCG@10 |
|---|---|---|---|---|---|---|
| MF | 0.0245±0.0009 | 0.0122±0.0003 | 0.0403±0.0026 | 0.0203±0.0009 | 0.0169±0.0008 | 0.0085±0.0005 |
| +(Ours) | 0.0637±0.0003 | 0.0318±0.0002 | 0.0928±0.0008 | 0.0460±0.0004 | 0.0500±0.0003 | 0.0250±0.0005 |
| GRU | 0.0752±0.0029 | 0.0405±0.0007 | 0.1157±0.0039 | 0.0623±0.0017 | 0.0560±0.0033 | 0.0303±0.0011 |
| +(Ours) | 0.1004±0.0013 | 0.0609±0.0002 | 0.1609±0.0028 | 0.0924±0.0030 | 0.0718±0.0030 | 0.0461±0.0015 |
| SASRec | 0.1492±0.0031 | 0.0712±0.0016 | 0.1792±0.0033 | 0.0850±0.0023 | 0.1350±0.0031 | 0.0646±0.0013 |
| +(Ours) | 0.1632±0.0039 | 0.0803±0.0022 | 0.2017±0.0048 | 0.0983±0.0043 | 0.1449±0.0040 | 0.0718±0.0014 |
| TiSASRec | 0.1545±0.0008 | 0.0734±0.0008 | 0.1872±0.0047 | 0.0884±0.0017 | 0.1391±0.0020 | 0.0663±0.0007 |
| +(Ours) | 0.1695±0.0014 | 0.0845±0.0007 | 0.1998±0.0035 | 0.1005±0.0015 | 0.1552±0.0010 | 0.0768±0.0006 |
| FEARec | 0.1570±0.0013 | 0.0741±0.0005 | 0.1893±0.0017 | 0.0891±0.0014 | 0.1417±0.0024 | 0.0670±0.0010 |
| +(Ours) | 0.1742±0.0007 | 0.0855±0.0004 | 0.2124±0.0010 | 0.1046±0.0013 | 0.1561±0.0007 | 0.0765±0.0003 |
| BSARec | 0.1613±0.0028 | 0.0755±0.0010 | 0.1933±0.0033 | 0.0906±0.0012 | 0.1455±0.0025 | 0.0680±0.0009 |
| +(Ours) | 0.1715±0.0009 | 0.0848±0.0003 | 0.2105±0.0039 | 0.1045±0.0014 | 0.1530±0.0012 | 0.0755±0.0004 |
Table 2-2: MovieLens
| Method | All Recall@10 | All NDCG@10 | Head Recall@10 | Head NDCG@10 | Tail Recall@10 | Tail NDCG@10 |
|---|---|---|---|---|---|---|
| MF | 0.0294±0.0009 | 0.0139±0.0004 | 0.0435±0.0022 | 0.0202±0.0011 | 0.0234±0.0004 | 0.0112±0.0003 |
| +(Ours) | 0.0375±0.0004 | 0.0177±0.0001 | 0.0728±0.0003 | 0.0341±0.0002 | 0.0225±0.0005 | 0.0108±0.0002 |
| GRU | 0.0803±0.0008 | 0.0388±0.0005 | 0.0970±0.0013 | 0.0471±0.0008 | 0.0732±0.0009 | 0.0353±0.0005 |
| +(Ours) | 0.0803±0.0009 | 0.0389±0.0003 | 0.1061±0.0005 | 0.0533±0.0004 | 0.0694±0.0012 | 0.0328±0.0006 |
| SASRec | 0.0839±0.0007 | 0.0359±0.0001 | 0.0966±0.0029 | 0.0416±0.0020 | 0.0785±0.0019 | 0.0335±0.0010 |
| +(Ours) | 0.1069±0.0007 | 0.0515±0.0003 | 0.1346±0.0008 | 0.0666±0.0005 | 0.0951±0.0009 | 0.0451±0.0005 |
| TiSASRec | 0.0921±0.0011 | 0.0386±0.0006 | 0.1059±0.0020 | 0.0447±0.0008 | 0.0863±0.0008 | 0.0361±0.0005 |
| +(Ours) | 0.1080±0.0009 | 0.0521±0.0005 | 0.1313±0.0024 | 0.0649±0.0013 | 0.0981±0.0010 | 0.0467±0.0004 |
| FEARec | 0.0861±0.0008 | 0.0361±0.0003 | 0.1014±0.0012 | 0.0427±0.0006 | 0.0797±0.0012 | 0.0333±0.0005 |
| +(Ours) | 0.0810±0.0030 | 0.0385±0.0016 | 0.0918±0.0053 | 0.0436±0.0028 | 0.0764±0.0021 | 0.0363±0.0011 |
| BSARec | 0.0727±0.0013 | 0.0303±0.0005 | 0.0827±0.0015 | 0.0348±0.0007 | 0.0696±0.0024 | 0.0287±0.0010 |
| +(Ours) | 0.0882±0.0007 | 0.0418±0.0002 | 0.1046±0.0014 | 0.0505±0.0008 | 0.0812±0.0006 | 0.0381±0.0001 |
Table 2-3: KuaiRand
| Method | All Recall@10 | All NDCG@10 | Head Recall@10 | Head NDCG@10 | Tail Recall@10 | Tail NDCG@10 |
|---|---|---|---|---|---|---|
| MF | 0.0393±0.0007 | 0.0190±0.0004 | 0.0468±0.0007 | 0.0227±0.0004 | 0.0114±0.0006 | 0.0050±0.0003 |
| +(Ours) | 0.0672±0.0002 | 0.0324±0.0001 | 0.0794±0.0001 | 0.0386±0.0002 | 0.0209±0.0011 | 0.0094±0.0003 |
| GRU | 0.0532±0.0006 | 0.0254±0.0004 | 0.0644±0.0010 | 0.0308±0.0006 | 0.0112±0.0010 | 0.0050±0.0005 |
| +(Ours) | 0.0684±0.0009 | 0.0336±0.0006 | 0.0860±0.0010 | 0.0424±0.0007 | 0.0017±0.0004 | 0.0006±0.0001 |
| SASRec | 0.0536±0.0009 | 0.0228±0.0003 | 0.0618±0.0013 | 0.0262±0.0005 | 0.0225±0.0016 | 0.0099±0.0008 |
| +(Ours) | 0.0735±0.0006 | 0.0358±0.0005 | 0.0921±0.0008 | 0.0449±0.0006 | 0.0036±0.0001 | 0.0013±0.0000 |
| TiSASRec | 0.0593±0.0010 | 0.0250±0.0004 | 0.0688±0.0013 | 0.0289±0.0005 | 0.0236±0.0004 | 0.0104±0.0001 |
| +(Ours) | 0.0749±0.0003 | 0.0363±0.0002 | 0.0939±0.0004 | 0.0456±0.0003 | 0.0036±0.0004 | 0.0013±0.0001 |
| FEARec | 0.0576±0.0016 | 0.0247±0.0006 | 0.0667±0.0024 | 0.0285±0.0011 | 0.0234±0.0018 | 0.0105±0.0011 |
| +(Ours) | 0.0611±0.0012 | 0.0305±0.0006 | 0.0770±0.0016 | 0.0385±0.0007 | 0.0008±0.0004 | 0.0003±0.0001 |
| BSARec | 0.0414±0.0005 | 0.0169±0.0002 | 0.0483±0.0012 | 0.0195±0.0004 | 0.0202±0.0009 | 0.0086±0.0004 |
| +(Ours) | 0.0679±0.0004 | 0.0333±0.0003 | 0.0854±0.0004 | 0.0421±0.0004 | 0.0019±0.0003 | 0.0007±0.0001 |
Table 3: Uniformity
| Backbone | Micro Video (Vanilla) | Micro Video (Ours) | MovieLens-1M (Vanilla) | MovieLens-1M (Ours) | KuaiRand (Vanilla) | KuaiRand (Ours) |
|---|---|---|---|---|---|---|
| MF | −3.9336±0.0025 | −3.8220±0.0035 | −3.6788±0.0014 | −3.7804±0.0027 | −3.7638±0.0037 | −3.7123±0.0052 |
| GRU | −3.0772±0.0286 | −3.8029±0.0031 | −2.6302±0.0183 | −3.5521±0.0185 | −2.5160±0.0451 | −3.1821±0.0073 |
| SASRec | −3.7316±0.0253 | −3.5652±0.0171 | −2.9970±0.0266 | −3.0466±0.0066 | −2.8904±0.0223 | −3.0597±0.0185 |
| TiSASRec | −3.6769±0.0101 | −3.6116±0.0111 | −2.7048±0.0217 | −2.9562±0.0246 | −2.5738±0.0286 | −2.8812±0.0194 |
| FEARec | −3.8309±0.0048 | −3.6014±0.0247 | −2.8215±0.0198 | −3.1607±0.0504 | −2.4473±0.0645 | −2.7468±0.0397 |
| BSARec | −3.8369±0.0093 | −3.6024±0.0136 | −2.9827±0.0257 | −3.2671±0.0137 | −2.4649±0.0506 | −3.0626±0.0120 |
We sincerely thank Reviewer ndQy for the careful examination of our experimental methodology and results, which has motivated additional experiments and strengthened our empirical evaluation.
W1. Correction of the Reporting Error in Table 1
We thank the reviewer for carefully identifying this error. The duplicated values in Table 1 resulted from a copy-and-paste error in the Vanilla SASRec row for Micro-Video. We have corrected the results in the revised manuscript, but the overall interpretation of the results remains unchanged. Please refer to Table 1 below or line 390 of the revised manuscript.
Table 1: Micro Video - SASRec Performance
W2. Clarification of Training Controls and Additional Loss-Controlled Experiments
We first clarify that the Vanilla backbones were trained using BCE with 15 negative samples per positive interaction. Thus, the number of negative samples was controlled when comparing Vanilla and Ours. We thank the reviewer for pointing out the need for clarification and have added this detail to the revised manuscript (revised manuscript, lines 1005-1007).
Following the reviewer's suggestion, we conducted an additional experiment using the same sampled-softmax loss as Ours, but with uniform popularity (λv(t;Φ)=1) and no centering penalty (γ=0). As shown in Tables 2-1, 2-2, 2-3, and 3 below (see also lines 1134–1201 of the revised manuscript), replacing BCE with the sampled-softmax loss substantially improves recommendation performance and representation uniformity. Nevertheless, Ours achieves further improvements, outperforming this loss-controlled baseline in 16 of 18 settings for overall Recall@10, all 18 settings for overall NDCG@10, and 12 of 18 settings for representation uniformity.
These results demonstrate that the improvements cannot be attributed solely to the change in the training objective, supporting the additional benefits of explicitly modeling temporal popularity and separating it from user-specific utility.
W3. Clarification Request Regarding the Theory–Practice Disconnect
We would appreciate further clarification on the specific disconnect between the theory and its practical implementation.
W4. Understanding the Role of the Centering Penalty in Practice
We thank the reviewer for this observation. To explain the strong performance at γ=0, we additionally examined whether the learned utility approximately satisfies the centering condition even without explicit regularization. However, we have not yet established why this centering phenomenon occurs. What is clear is that identifiability is not guaranteed unless the centering constraint is satisfied. Therefore, we believe that introducing the centering penalty is necessary to prevent cases in which the observed centering phenomenon does not arise, thereby encouraging the proposed model to satisfy the condition required for identifiability. We have added this analysis and the corresponding results (revised manuscripts, lines 469--494).
W5.
We are currently conducting additional experiments to address weakness 5. We will post our response as soon as the experiments are completed.