- 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
- 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.
- 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.
- 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
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
| Model | HR@5 | NDCG@5 | MRR@5 | HR@10 | NDCG@10 | MRR@10 | HR@20 | NDCG@20 | MRR@20 |
|---|---|---|---|---|---|---|---|---|---|
| SASRec | 0.0350 | 0.0228 | 0.0188 | 0.0550 | 0.0292 | 0.0214 | 0.0879 | 0.0374 | 0.0236 |
| NextItNet | 0.0347 | 0.0215 | 0.0171 | 0.0540 | 0.0276 | 0.0196 | 0.0852 | 0.0354 | 0.0217 |
| STOSA | 0.0346 | 0.0239 | 0.0204 | 0.0534 | 0.0300 | 0.0229 | 0.0832 | 0.0375 | 0.0250 |
| SR-PLR | 0.0357 | 0.0267 | 0.0237 | 0.0538 | 0.0325 | 0.0261 | 0.0852 | 0.0403 | 0.0282 |
| Attr-SASRec | 0.0343 | 0.0205 | 0.0160 | 0.0571 | 0.0279 | 0.0191 | 0.0900 | 0.0362 | 0.0213 |
| Attr-NextItNet | 0.0322 | 0.0188 | 0.0145 | 0.0576 | 0.0270 | 0.0179 | 0.0897 | 0.0351 | 0.0201 |
| Attr-SR-PLR | 0.0365 | 0.0230 | 0.0186 | 0.0589 | 0.0302 | 0.0215 | 0.0914 | 0.0384 | 0.0237 |
| FDSA | 0.0397 | 0.0256 | 0.0210 | 0.0583 | 0.0316 | 0.0234 | 0.0900 | 0.0395 | 0.0256 |
| DIF-SR | 0.0307 | 0.0195 | 0.0158 | 0.0522 | 0.0263 | 0.0186 | 0.0864 | 0.0349 | 0.0209 |
| NCR | 0.0292 | 0.0180 | 0.0144 | 0.0510 | 0.0250 | 0.0172 | 0.0818 | 0.0327 | 0.0193 |
| Mamba4Rec | 0.0469 | 0.0301 | 0.0246 | 0.0737 | 0.0387 | 0.0281 | 0.1126 | 0.0484 | 0.0307 |
| DIFF | 0.0450 | 0.0301 | 0.0252 | 0.0768 | 0.0404 | 0.0295 | 0.1202 | 0.0513 | 0.0325 |
| TAPR | 0.0528 | 0.0358 | 0.0302 | 0.0823 | 0.0452 | 0.0341 | 0.1243 | 0.0558 | 0.0370 |
Yelp
| Model | HR@5 | NDCG@5 | MRR@5 | HR@10 | NDCG@10 | MRR@10 | HR@20 | NDCG@20 | MRR@20 |
|---|---|---|---|---|---|---|---|---|---|
| SASRec | 0.0310 | 0.0204 | 0.0170 | 0.0504 | 0.0266 | 0.0195 | 0.0847 | 0.0352 | 0.0218 |
| NextItNet | 0.0310 | 0.0209 | 0.0175 | 0.0513 | 0.0274 | 0.0202 | 0.0858 | 0.0360 | 0.0225 |
| STOSA | 0.0389 | 0.0261 | 0.0219 | 0.0604 | 0.0330 | 0.0247 | 0.0951 | 0.0417 | 0.0271 |
| SR-PLR | 0.0343 | 0.0224 | 0.0185 | 0.0567 | 0.0295 | 0.0214 | 0.0930 | 0.0387 | 0.0239 |
| Attr-SASRec | 0.0224 | 0.0137 | 0.0108 | 0.0405 | 0.0195 | 0.0132 | 0.0678 | 0.0263 | 0.0150 |
| Attr-NextItNet | 0.0319 | 0.0210 | 0.0174 | 0.0539 | 0.0280 | 0.0203 | 0.0874 | 0.0364 | 0.0225 |
| Attr-SR-PLR | 0.0229 | 0.0145 | 0.0117 | 0.0396 | 0.0198 | 0.0139 | 0.0705 | 0.0276 | 0.0160 |
| FDSA | 0.0226 | 0.0143 | 0.0116 | 0.0395 | 0.0197 | 0.0138 | 0.0675 | 0.0267 | 0.0157 |
| DIF-SR | 0.0455 | 0.0314 | 0.0268 | 0.0721 | 0.0399 | 0.0302 | 0.1132 | 0.0502 | 0.0330 |
| NCR | 0.0207 | 0.0129 | 0.0103 | 0.0382 | 0.0185 | 0.0126 | 0.0643 | 0.0250 | 0.0144 |
| Mamba4Rec | 0.0342 | 0.0208 | 0.0165 | 0.0595 | 0.0289 | 0.0197 | 0.1006 | 0.0392 | 0.0225 |
| DIFF | 0.0414 | 0.0265 | 0.0216 | 0.0701 | 0.0357 | 0.0253 | 0.1153 | 0.0470 | 0.0284 |
| TAPR | 0.0461 | 0.0305 | 0.0254 | 0.0760 | 0.0401 | 0.0293 | 0.1222 | 0.0517 | 0.0324 |
H&M
| Model | HR@5 | NDCG@5 | MRR@5 | HR@10 | NDCG@10 | MRR@10 | HR@20 | NDCG@20 | MRR@20 |
|---|---|---|---|---|---|---|---|---|---|
| SASRec | 0.0500 | 0.0332 | 0.0277 | 0.0791 | 0.0426 | 0.0315 | 0.1211 | 0.0531 | 0.0344 |
| NextItNet | 0.0485 | 0.0324 | 0.0272 | 0.0779 | 0.0418 | 0.0310 | 0.1169 | 0.0516 | 0.0336 |
| STOSA | 0.0633 | 0.0451 | 0.0391 | 0.0914 | 0.0540 | 0.0427 | 0.1321 | 0.0643 | 0.0455 |
| SR-PLR | 0.0509 | 0.0339 | 0.0284 | 0.0808 | 0.0436 | 0.0323 | 0.1254 | 0.0548 | 0.0354 |
| Attr-SASRec | 0.0443 | 0.0285 | 0.0234 | 0.0714 | 0.0373 | 0.0270 | 0.1137 | 0.0479 | 0.0299 |
| Attr-NextItNet | 0.0475 | 0.0309 | 0.0254 | 0.0766 | 0.0402 | 0.0292 | 0.1183 | 0.0507 | 0.0321 |
| Attr-SR-PLR | 0.0458 | 0.0288 | 0.0232 | 0.0753 | 0.0382 | 0.0271 | 0.1200 | 0.0494 | 0.0301 |
| FDSA | 0.0374 | 0.0234 | 0.0189 | 0.0631 | 0.0317 | 0.0223 | 0.1038 | 0.0419 | 0.0250 |
| DIF-SR | 0.0695 | 0.0477 | 0.0406 | 0.1057 | 0.0593 | 0.0453 | 0.1515 | 0.0708 | 0.0484 |
| NCR | 0.0225 | 0.0140 | 0.0112 | 0.0400 | 0.0196 | 0.0135 | 0.0660 | 0.0261 | 0.0153 |
| Mamba4Rec | 0.0636 | 0.0418 | 0.0347 | 0.0974 | 0.0527 | 0.0391 | 0.1464 | 0.0650 | 0.0425 |
| DIFF | 0.0706 | 0.0490 | 0.0419 | 0.1035 | 0.0596 | 0.0462 | 0.1498 | 0.0712 | 0.0494 |
| TAPR | 0.0755 | 0.0514 | 0.0435 | 0.1120 | 0.0631 | 0.0483 | 0.1635 | 0.0760 | 0.0518 |
TAPR achieves the best results on most metrics across the three datasets.
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.
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
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
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.