acceptodds
Under review as a conference paper at ICLR 2027

Attribute-Structured Temporal Rule Representations for Sequential Recommendation

Abstract

Logic-based sequential recommenders introduce logical operators to impose structured reasoning for better recommendation, but few of them explicitly explore the impact of logical structure on this complex preference learning. Item attributes offer a natural grounding for making the role of logical structure explicit and empirically testable. yet existing methods rarely construct logic at this level. To address this gap, we introduce TAPR, a Temporal Attribute Probabilistic Rule model that grounds logical preference rules in explicit item attributes derived from a user’s interaction history. Probabilistic conjunction then composes this evidence into a structured rule that directly shapes the final ranking score. Implemented in a Beta representation space, this formulation captures soft preference evidence while remaining closed under probabilistic conjunction, allowing composed rules to stay in the same representation space for ranking. Experiments on three real-world datasets show that TAPR is competitive with strong sequential recommendation baselines. Synthetic experiments further verify that the explicit logical structure contributes to ranking behavior rather than serving only as a descriptive representation. Additional empirical analyses show that the learned rules are behaviorally faithful and that their utility depends on data availability and attribute-combination difficulty.

Primary area:
interpretability and explainable AI
Keywords:
Sequential Recommendation, Temporal Rule Induction, Attribute-Aware Recommendation, Probabilistic Rules, Beta Distributions, Explainable Recommendation

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_s6w7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TAPR, a sequential recommendation model that grounds item identities and attributes in historical interaction events. These pieces of evidence are represented as Beta distributions and composed into a history-level user rule using weights conditioned on a local causal CNN, same-type relations, and ordered inter-event relations. Candidates are ranked by their KL divergence to the rule distribution. The evaluation includes three real-world datasets, synthetic interventions, rule-faithfulness tests, and analyses under limited training data.

Critical strengths
  1. Preserving attributes as event-bound evidence allows the model to distinguish contributions from the same attribute type across different interactions. The evidence-to-rule-to-ranking computation is clearly defined and supports targeted structural interventions.
  2. Under full-catalog evaluation, TAPR achieves a substantial improvement on MovieLens and gains on H&M. The paper appropriately acknowledges the smaller improvement on Yelp rather than presenting the advantage as equally strong across datasets.
  3. The paper distinguishes comparisons between separately trained variants from branch interventions in a frozen model, evaluates extracted rules against matched controls, and reports both inconsistent gains from relation branches and faithfulness failures. These analyses help characterize which structures the model actually uses.
Critical weaknesses

1.The full model outperforms CNN-only on MovieLens in Table 3, but setting the relational residual coefficient to zero gives a slightly higher mean in Table 13. CNN-only also performs slightly better on Yelp, as the authors acknowledge. Please explain how these experimental settings differ, whether the observed differences are consistent across runs, and under what conditions the relation modules provide additional benefits. 2.The paper states that it removes the Beta-based rule representation, but does not clearly explain how this variant represents and combines historical evidence or scores candidate items. Please specify these choices and which components remain unchanged, so that readers can assess whether the comparison isolates the contribution of the Beta-based rule representation. 3.The experimental evaluation only adopts two metrics, HR@10 and NDCG@10, which are somewhat insufficient. It would be beneficial to incorporate additional metrics to provide a more comprehensive assessment of the model’s performance.

Questions for authors

1.Could the authors clarify how CNN-only in Table 3 differs from the zero-relational-residual setting in Table 13, and explain why CNN-only performs slightly better than the full model on Yelp? 2.How does strict-no-rule represent and combine historical evidence and score candidate items after removing the Beta-based rule representation? Which components and training settings remain unchanged relative to the full model? 3.Please clarify the intended interpretation of a “user rule”: is it a model representation for predicting observed interactions, or is it intended to reflect the user’s actual choice mechanism? If the latter interpretation is intended, how do sampled non-interacted negatives and potential exposure bias limit that interpretation?

5 comments
AuthorsOpenReview ↗

Response to Reviewer s6w7(1/3)

Q1: How does CNN-only in Table 3 differ from the zero-relational-residual setting in Table 13, and why does CNN-only performs slightly better than the full model on Yelp?

A1. We thank the reviewer for pointing out the discrepancy between Tables 3 and 13. CNN-only and the zero-residual setting implement the same ablation. The submitted Table 3 used the five-seed MovieLens ablation setup, but its CNN-only entry inadvertently retained the result from earlier runs with an early-stopping patience of 10 epochs. Rerunning CNN-only with the updated Table 3 configuration (an early-stopping patience of 50 epochs) yields 0.0435 NDCG@10. Table 13 reports 0.0473 from a separate three-seed sensitivity experiment. We apologize for the reporting error and have corrected the MovieLens CNN-only result in the revised Table 3.

We expect cross-event relations to help when an attribute value recurs after a gap, since the earlier occurrence lies outside the immediate context. On Yelp, the local CNN alone provides competitive evidence weights for the Beta rule, leaving the benefit of the relation module unclear. To examine when relation-based weighting improves ranking beyond local causal encoding, we compared TAPR with CNN-only on an exploratory slice defined from the input history. In the slice, an attribute value in the latest interaction also appears earlier in the history and is absent from the five immediately preceding interactions.

Table A. NDCG@10 overall and in an attribute recurrence slice

DatasetEvaluation groupUsersTAPRCNN-only
YelpOverall10,0000.04010.0408
YelpCoffee & Tea recurrence2230.02400.0173

Results in Table A show that TAPR outperformed CNN-only in the exploratory recurrence slice, suggesting that cross-event weighting can help when evidence for the same attribute is spread across separated interactions.

Q2: How does strict-no-rule represent and combine historical evidence and score candidate items after removing the Beta-based rule representation? Which components and training settings remain unchanged relative to the full model?

A2. We thank the reviewer for asking how strict-no-rule represents the history and scores candidates. Strict-no-rule represents item IDs and attribute values as point embeddings. It pools each item's attribute embeddings and combines them with its ID embedding. A causal CNN encodes the resulting item sequence, and its output at the last observed interaction serves as the history representation. Candidates use the same item encoder and are ranked by dot product with this representation. Both models use the same item IDs and attributes, the sampled-negative BPR objective, and full-catalog evaluation.

Strict-no-rule evaluates the Beta rule design as a whole. To examine this core design more closely, we compare TAPR with two additional variants. Continuous-dot retains TAPR's causal CNN and relation branches for evidence weighting and uses the same weighted-sum aggregation, but replaces the Beta parameterization with signed vectors and KL matching with scaled dot-product scoring. We use this variant to test whether weighted historical evidence can support comparable ranking without requiring the pooled representation to define a Beta rule. Beta-log-L2 retains TAPR's Beta rule construction but replaces KL matching with the negative scaled squared Euclidean distance between the log-transformed Beta parameters of the user rule and each candidate. We use this variant to test whether KL matching improves ranking when the Beta rule is constructed in the same way.

Table B. Ranking performance on MovieLens

ModelNDCG@10HR@10
TAPR0.04520.0823
Continuous-dot0.02970.0571
Beta-log-L20.03210.0587

On MovieLens, TAPR outperforms Continuous-dot on both metrics. With the same Beta rule construction, KL matching also outperforms log-parameter distance.

Q3: Is a “user rule” a model representation for predicting observed interactions, or does it reflect the user’s actual choice mechanism? If the latter, how do sampled non-interacted negatives and potential exposure bias limit that interpretation?

A3. We thank the reviewer for raising this distinction. We intend the first interpretation: the user rule is TAPR’s representation of the observed history for predicting the next interaction. In Section 4.4, we test whether retaining the extracted evidence preserves TAPR’s top-ranked item and whether removing it changes the ranking. Across all three datasets, the extracted rules outperform matched controls on all three faithfulness metrics (Figure 3), supporting their use in explaining TAPR’s ranking decisions.

AuthorsOpenReview ↗

Response to Reviewer s6w7(2/3)

Q4: Incorporate additional metrics to provide a more comprehensive assessment of the model’s performance

A4. We thank the reviewer for suggesting additional ranking metrics. We report mean HR, NDCG, and MRR at cutoffs of 5, 10, and 20. Following suggestions from other reviewers, we also evaluated Mamba4Rec [1] and DIFF [2]. We have added both baselines to Table 2 in the revised manuscript.

Table C. Complete overall performance

MovieLens

ModelHR@5NDCG@5MRR@5HR@10NDCG@10MRR@10HR@20NDCG@20MRR@20
SASRec0.03500.02280.01880.05500.02920.02140.08790.03740.0236
NextItNet0.03470.02150.01710.05400.02760.01960.08520.03540.0217
STOSA0.03460.02390.02040.05340.03000.02290.08320.03750.0250
SR-PLR0.03570.02670.02370.05380.03250.02610.08520.04030.0282
Attr-SASRec0.03430.02050.01600.05710.02790.01910.09000.03620.0213
Attr-NextItNet0.03220.01880.01450.05760.02700.01790.08970.03510.0201
Attr-SR-PLR0.03650.02300.01860.05890.03020.02150.09140.03840.0237
FDSA0.03970.02560.02100.05830.03160.02340.09000.03950.0256
DIF-SR0.03070.01950.01580.05220.02630.01860.08640.03490.0209
NCR0.02920.01800.01440.05100.02500.01720.08180.03270.0193
Mamba4Rec0.04690.03010.02460.07370.03870.02810.11260.04840.0307
DIFF0.04500.03010.02520.07680.04040.02950.12020.05130.0325
TAPR0.05280.03580.03020.08230.04520.03410.12430.05580.0370

Yelp

ModelHR@5NDCG@5MRR@5HR@10NDCG@10MRR@10HR@20NDCG@20MRR@20
SASRec0.03100.02040.01700.05040.02660.01950.08470.03520.0218
NextItNet0.03100.02090.01750.05130.02740.02020.08580.03600.0225
STOSA0.03890.02610.02190.06040.03300.02470.09510.04170.0271
SR-PLR0.03430.02240.01850.05670.02950.02140.09300.03870.0239
Attr-SASRec0.02240.01370.01080.04050.01950.01320.06780.02630.0150
Attr-NextItNet0.03190.02100.01740.05390.02800.02030.08740.03640.0225
Attr-SR-PLR0.02290.01450.01170.03960.01980.01390.07050.02760.0160
FDSA0.02260.01430.01160.03950.01970.01380.06750.02670.0157
DIF-SR0.04550.03140.02680.07210.03990.03020.11320.05020.0330
NCR0.02070.01290.01030.03820.01850.01260.06430.02500.0144
Mamba4Rec0.03420.02080.01650.05950.02890.01970.10060.03920.0225
DIFF0.04140.02650.02160.07010.03570.02530.11530.04700.0284
TAPR0.04610.03050.02540.07600.04010.02930.12220.05170.0324

H&M

ModelHR@5NDCG@5MRR@5HR@10NDCG@10MRR@10HR@20NDCG@20MRR@20
SASRec0.05000.03320.02770.07910.04260.03150.12110.05310.0344
NextItNet0.04850.03240.02720.07790.04180.03100.11690.05160.0336
STOSA0.06330.04510.03910.09140.05400.04270.13210.06430.0455
SR-PLR0.05090.03390.02840.08080.04360.03230.12540.05480.0354
Attr-SASRec0.04430.02850.02340.07140.03730.02700.11370.04790.0299
Attr-NextItNet0.04750.03090.02540.07660.04020.02920.11830.05070.0321
Attr-SR-PLR0.04580.02880.02320.07530.03820.02710.12000.04940.0301
FDSA0.03740.02340.01890.06310.03170.02230.10380.04190.0250
DIF-SR0.06950.04770.04060.10570.05930.04530.15150.07080.0484
NCR0.02250.01400.01120.04000.01960.01350.06600.02610.0153
Mamba4Rec0.06360.04180.03470.09740.05270.03910.14640.06500.0425
DIFF0.07060.04900.04190.10350.05960.04620.14980.07120.0494
TAPR0.07550.05140.04350.11200.06310.04830.16350.07600.0518

TAPR achieves the best results on most metrics across the three datasets.

AuthorsOpenReview ↗

Response to Reviewer s6w7(3/3)

References

[1] Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee. Mamba4Rec: Towards Efficient Sequential Recommendation with Selective State Space Models.

[2] Hye-young Kim, Minjin Choi, Sunkyung Lee, Ilwoong Baek, and Jongwuk Lee. DIFF: Dual Side-Information Filtering and Fusion for Sequential Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’25), pp. 1624–1633, 2025.

Reviewer_6VUGOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TAPR, a sequential recommender that models user preferences as probabilistic rules over item attributes. It encodes attributes as Beta distributions (following BetaE) and combines them using a product-of-experts conjunction. Temporal relations are modeled with a causal CNN and same-type and THEN channels. Candidates are ranked by negative KL divergence to the user rule. A key strength is that the paper tests whether the rule structure actually affects ranking. It uses synthetic interventions and matched-control faithfulness metrics. The gains are large on MovieLens (+39.1% NDCG) but modest on Yelp (+0.5%) and H&M (+6.4%).

Critical strengths

The validation looks good. The interventions change only semantics or timing while keeping the history, positions, and catalog fixed. The faithfulness metrics also use matched controls and bootstrap confidence intervals. This goes beyond simply claiming that the interpretable structure drives predictions.

The formalization is also solid. It proves closure of the conjunction and clarifies that it represents neither Boolean semantics nor calibrated probabilities. The authors also acknowledge that gains are concentrated on MovieLens. They note that the rule representation receives more consistent support than the relation channels.

Critical weaknesses

The baselines are somewhat outdated, which makes the reported gains difficult to assess. The newest appear to be SR-PLR (2023) and DIF-SR (2022), while newer methods such as Mamba4Rec and LRURec (both 2024), as well as recent side-information fusion methods, are missing. It is therefore unclear whether the reported 39.1% gain on MovieLens would hold against stronger, more recent baselines.

The contribution of the explicit relation channels remains unclear. CNN-only performs slightly better on Yelp, and setting λrel=0 improves MovieLens NDCG (0.0473 vs. 0.0463). While synthetic interventions support the role of THEN, its benefit on real-world data remains uncertain. Overall, the rule representation receives more consistent support than the explicit relation channels.

These concerns are addressable. I would reconsider if the rebuttal included stronger 2024–2026 baselines and clarified when the temporal channels help on real-world data.

Questions for authors

How does TAPR compare with Mamba4Rec and a recent side-information fusion method under the same evaluation setup? This would help clarify whether its gains hold against stronger, more recent baselines.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6VUG

Q1: How does TAPR compare with Mamba4Rec and a recent side-information fusion method under the same evaluation setup?

A1. We thank the reviewer for suggesting recent baselines. We evaluated Mamba4Rec [1] and DIFF [2] as additional baselines. These comparisons use the same data splits and full-catalog evaluation protocol, with HR@10 and NDCG@10 as in the main paper.

Table A. Full-catalog comparison with Mamba4Rec and DIFF.

ModelMovieLens HR@10MovieLens NDCG@10Yelp HR@10Yelp NDCG@10H&M HR@10H&M NDCG@10
Mamba4Rec [1]0.07370.03870.05950.02890.09740.0527
DIFF [2]0.07680.04040.07010.03570.10350.0596
TAPR0.08230.04520.07600.04010.11200.0631

As shown in Table A, TAPR achieves the highest performance across all three datasets. DIFF is the stronger of the two added baselines on every dataset. Relative to DIFF, TAPR improves HR@10 and NDCG@10 by 7.1% and 11.8% on MovieLens, 8.4% and 12.4% on Yelp, and 8.2% and 6.0% on H&M, respectively. We have included these baselines and their results in Table 2 of the revised manuscript.

Q2: When do the temporal channels help on real-world data?

A2. We thank the reviewer for asking when the temporal channels help on real-world data. As described in Section 3.3, TAPR uses cross-event relations to adjust the weights assigned to historical evidence. To test whether this adjustment improves ranking beyond local causal encoding, we compare TAPR with CNN-only, which derives evidence weights from the causal CNN alone. We expect relation-based weighting to be particularly useful when the same attribute value recurs across separated interactions, where that recurrence is harder to detect from local context.

We define two slices on H&M based on each user’s interaction history. A user belongs to a slice if the latest interaction shares an attribute value with an earlier interaction and that value is absent from the five immediately preceding interactions. We apply this criterion to two attribute values: the "socks and tights" garment group and the colour blue.

Table B. Relation ablations on H&M slices (NDCG@10).

Evaluation groupUsersTAPRCNN-only
Overall10,0000.06310.0612
Socks/tights recurrence1080.05610.0324
Blue recurrence5030.07320.0614

TAPR’s NDCG@10 gains over CNN-only were 0.0237 on the socks/tights slice and 0.0118 on the blue slice, compared with 0.0019 overall. The larger gains in these two exploratory slices suggest that adjusting evidence weights using cross-event relations can be especially useful when the same attribute value appears in interactions separated by a gap.

References

[1] Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee. Mamba4Rec: Towards Efficient Sequential Recommendation with Selective State Space Models.

[2] Hye-young Kim, Minjin Choi, Sunkyung Lee, Ilwoong Baek, and Jongwuk Lee. DIFF: Dual Side-Information Filtering and Fusion for Sequential Recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’25), pp. 1624–1633, 2025.

Thanks for providing the additional information. The two added comparisons partially answer my first question. The H&M slices show when the channels help, but the overall gain is negligible meaning gains from typical attributions get canceled by others. Thus my second question is also partly answered. I will keep my rating in this case.

We sincerely thank you for your constructive feedback and your time in reviewing our work.

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

This paper focuses on predicting next-item preferences not just by looking at what users liked in the past, but by extracting explicit rules around which attributes they preferred and how those tastes evolved over time. While TAPR isn't the first to use item features or temporal dynamics, its key strength lies in explicitly preserving user attributes as distinct pieces of evidence and mapping out exactly how these relationships translate into the final recommendation score.

Critical strengths
  1. First, the mathematical formulation is solid and easy to follow, clearly laying out the Beta conjugate operations, the KL-divergence-based scoring, and the pairwise BPR loss.

  2. The experimental design is also well thought out, going beyond simple benchmark leaderboards. Strong ranking metrics alone don't prove that an architecture works as intended, and the authors address this directly with targeted validation tests—such as modifying specific attributes in synthetic data, permuting interaction sequences, and injecting irrelevant noise. In particular, ablating the THEN module within the exact same backbone serves as a clean, effective check to isolate its actual contribution.

Critical weaknesses
  1. The necessity of the core Temporal Relation module remains questionable. While the paper claims that explicitly modeling Same-type and THEN relations captures preference dynamics better than existing sequential models, the ablation study in Section 4.3 shows mixed results: TAPR yields noticeable gains on MovieLens, but the improvement is marginal on H&M, and a simple CNN-only variant actually outperforms it on Yelp.

  2. On MovieLens, even stripped-down variants of TAPR (lacking its core components) consistently beat the strongest baseline (SR-PLR). This makes it difficult to pinpoint where the performance gains actually come from, raising concerns about whether the baseline was properly tuned or sufficiently strong.

Questions for authors
  1. What is the actual, tangible contribution of the temporal relation module? The empirical results leave this unclear, and I would appreciate the authors' clarification on this point.

  2. The final user rule appears to be computed as a simplex-weighted average of Beta parameters. What distinct advantages does this specific formulation offer over standard attention pooling or other continuous aggregation baselines applied to the same evidence representations?

4 comments
AuthorsOpenReview ↗

Response to Reviewer CDGc (1/2)

Q1: What is the actual contribution of the temporal relation module?

A1. We thank the reviewer for asking us to clarify the contribution of the temporal relation module. To test whether cross-event relations improve ranking by adjusting the weights of historical evidence, we compare TAPR with CNN-only, which derives these weights from the causal CNN alone. We expect relation-based weighting to add the most value when an attribute recurs across distant interactions, where its recurrence is less apparent from local context. To examine this on H&M, we selected histories in which the latest interaction shares an attribute value with an earlier interaction and the preceding five interactions lack that value. We report two exploratory slices, one for the “socks and tights” garment group and one for the colour blue.

Table A. Relation ablations on H&M slices (NDCG@10).

Evaluation groupUsersTAPRCNN-only
Overall10,0000.06310.0612
Socks/tights recurrence1080.05610.0324
Blue recurrence5030.07320.0614

TAPR’s gains over CNN-only are larger in both recurrence slices than in the overall evaluation. This suggests that using relations between separated interactions to adjust evidence weights can improve ranking in these histories. Figure 2(b) also provides evidence for the role of THEN: disabling it at inference reduced the bidirectional oracle margin by 0.5930 on synthetic pairs where event order determined the correct target.

CNN-only removes the explicit relation channels but still represents item IDs and attributes as event-bound Beta evidence, composes that evidence into a user rule using weights from the causal CNN, and ranks candidates by KL divergence. Its performance against SR-PLR shows that the MovieLens gain cannot be attributed to the relation channels alone. To address the concern about baseline strength, we also evaluated Mamba4Rec [1] and DIFF [2] using the same data splits and evaluation protocol. They achieved 0.0387 and 0.0404 NDCG@10, respectively, compared with 0.0452 for TAPR, an 11.8% improvement over DIFF.

AuthorsOpenReview ↗

Response to Reviewer CDGc (2/2)

Q2: What advantages does the Beta conjunction offer over standard attention pooling or continuous aggregation?

A2. We thank the reviewer for pointing out the connection to attention pooling. As shown in Appendix B (Corollary 1), the normalized weighted product of Beta densities yields the same Beta rule as averaging each parameter separately with the same simplex weights. We call this operation a probabilistic conjunction because it combines the support that different pieces of evidence assign to the same latent preference state, forming a soft intersection (Appendix B.3). In TAPR, local context and cross-event relations determine how much each typed, position-bound attribute observation contributes to the user rule. This gives rule construction and ranking a shared interpretation. The conjunction defines what the historical evidence jointly supports, and negative KL scores how closely each candidate's Beta distribution matches that rule.

PTSR [3] provides a precedent for this design in sequential recommendation. It uses probabilistic conjunction to compose items within contiguous windows into patterns, then uses KL divergence to assess how closely each candidate matches those patterns. TAPR applies the same composition-and-matching principle to a different unit of evidence. It combines typed attribute observations across historical events into a user rule, with cross-event relations helping determine their weights.

To assess the ranking benefit of composing Beta evidence and matching the resulting rule with KL divergence, we compare TAPR with two variants on three real-world datasets. Continuous-dot encodes the same attribute inputs as continuous vectors in place of Beta parameterization, retains TAPR's weight-generation architecture, and ranks candidates by scaled dot product after weighted pooling. Beta-log-L2 retains the Beta rule construction and replaces KL divergence with scaled squared Euclidean distance between log-transformed Beta parameters.

Table B. Ranking performance across three datasets

DatasetModelNDCG@10HR@10
MovieLensTAPR0.04520.0823
MovieLensContinuous-dot0.02970.0571
MovieLensBeta-log-L20.03210.0587
YelpTAPR0.04010.0760
YelpContinuous-dot0.02290.0463
YelpBeta-log-L20.03560.0693
H&MTAPR0.06310.1120
H&MContinuous-dot0.02810.0573
H&MBeta-log-L20.05860.1062

As shown in Table B, TAPR outperforms both variants on NDCG@10 and HR@10, supporting a unified rule-based design in which probabilistic conjunction builds the user rule from historical evidence and KL matching uses that same rule to rank candidates.

References

[1] Chengkai Liu, Jianghao Lin, Jianling Wang, Hanzhou Liu, and James Caverlee. Mamba4Rec: Towards Efficient Sequential Recommendation with Selective State Space Models. Presented at the Resource-Efficient Learning for Knowledge Discovery Workshop (RelKD), KDD 2024.

[2] Hye-young Kim, Minjin Choi, Sunkyung Lee, Ilwoong Baek, and Jongwuk Lee. DIFF: Dual Side-Information Filtering and Fusion for Sequential Recommendation. SIGIR 2025, pp. 1624–1633.

[3] Kun Ma, Cong Xu, Zeyuan Chen, and Wei Zhang. Pattern-Wise Transparent Sequential Recommendation. IEEE Transactions on Knowledge and Data Engineering, 37(6), pp. 3627–3640, 2025.

Thank you for the authors’ response. The additional analysis and experiments provided by the authors have addressed my main concerns, and I am willing to raise my score to 3.

We sincerely thank you for your constructive feedback and your time in reviewing our work. Your review helps us evaluate our method more thoroughly.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their careful evaluations and constructive discussion. Their feedback has helped us clarify the contribution of TAPR, strengthen its empirical support, and better characterize its applicability. Below, we summarize the main points addressed during the discussion.

1. Contribution of the Beta rule and matching function (Reviewers CDGc and s6w7). We clarified why the aggregation is described as Beta conjunction: it combines evidence through a normalized weighted product of densities, whose closed form under simplex weights is a weighted average of Beta parameters. We also detailed the strict-no-rule implementation and added Continuous-dot and Beta-log-L2 comparisons. These comparisons support the effectiveness of TAPR’s Beta rule construction and KL-based matching.

2. Contribution and applicability of temporal relations (All reviewers). We added exploratory recurrence-slice analyses to examine when relation-based weighting improves ranking beyond CNN-only. These slices contain histories in which an attribute value in the latest interaction also appears earlier but is absent from the five immediately preceding interactions. TAPR outperforms CNN-only on the examined recurrence slices, suggesting that cross-event relations can help when evidence for the same attribute value is spread across separated interactions.

3. Recent baselines and broader ranking evaluation (Reviewers 6VUG and s6w7). In response to 6VUG, we evaluated Mamba4Rec and DIFF using the same data splits and full-catalog evaluation protocol. TAPR outperforms both additional baselines on HR@10 and NDCG@10 across all three datasets. These results and the accompanying discussion have been incorporated into Table 2. In response to s6w7, we additionally reported HR, NDCG, and MRR at cutoffs of 5, 10, and 20. TAPR achieves the best results on most metrics across the three datasets.

We once again thank the reviewers for their insightful comments and constructive suggestions. TAPR provides a framework for composing event-bound attribute evidence into explicit probabilistic preference rules for sequential recommendation. Local context and cross-event relations determine how historical evidence contributes to a Beta rule, which directly guides candidate ranking through KL matching. Our results support competitive recommendation performance, while the explicit evidence-to-rule-to-ranking path enables targeted interventions and faithfulness analysis to examine how the learned structure participates in ranking decisions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.