acceptodds
Under review as a conference paper at ICLR 2027

SL4SR: Selective Learning for Sequential Recommendation

Abstract

Recent advances in deep learning have substantially improved sequential recommendation (SR) by enabling more expressive architectures to capture complex user behavior patterns. However, the increased model capacity can also make SR models more prone to fitting incidental, transient, or weakly informative interactions that are less generalizable to future user preferences. This issue is further amplified by the prevailing training paradigm in SR, where standard objectives typically treat all training instances uniformly, regardless of their usefulness for improving the current model. To address this limitation, we propose , a simple and effective elective earning framework for equential ecommendation. SL4SR leverages a pre-trained estimator as a reference model and compares its sample-wise loss with that of the current backbone to estimate the learning utility of each training instance. Based on this estimator-guided score, the model selectively updates on informative high-score instances while filtering out instances that are less likely to improve generalization. The proposed framework is model-agnostic and introduces no additional inference overhead. Extensive experiments on six real-world datasets with four representative SR backbones demonstrate the effectiveness of SL4SR, yielding average improvements of 16.97%, 17.65%, 17.07%, and 4.85% for SASRec, GRU4Rec, FMLPRec, and Mamba4Rec, respectively.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Sequential Recommendation, Selective Learning, Sample Selection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_eR5iOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper highlights the limitation of uniformly optimizing all training instances in sequential recommendation and proposes a selective learning framework called SL4SR for resolving this problem. SL4SR first pre-trains a reference model on the full training set. It then estimates the relative learnability of each training instance using the loss difference between the current backbone and the reference model, and selectively updates the backbone using only high-scoring instances within each mini-batch. Experiments on several public datasets report consistent performance improvements over the reproduced baselines.

Critical strengths
  1. The paper is clearly written and well organized. The motivation and proposed method are easy to understand.
  2. The experiments cover multiple real-world datasets and different backbone architectures. The paper also provides a relatively comprehensive set of ablation studies and additional analyses to investigate the behavior of the proposed method.
Critical weaknesses
  1. SL4SR is highly similar to RHO-LOSS and seems to be a straightforward adaptation in sequential recommendation with limited methodological novelty. Compared to RHO-LOSS, SL4SR replaces the out-of-sample reference loss with the in-sample reference loss, but the motivation of this change are not sufficiently discussed or experimentally validated. Besides, the experimental configuration of the RHO-LOSS baseline is not described in sufficient detail. Important settings (e.g., the held-out subset ratio, reference model configuration, and sample selection ratio) are missing, making it difficult to assess the fairness and reproducibility of the comparison.
  2. The reported performance of several backbone models is substantially lower than previously published results. For example, compared with the full-ranking results of the original FMLPRec paper, the HR@10 results of SASRec and FMLPRec in SL4SR are consistently lower and even SL-enhanced models do not reach the previously reported performance.
Questions for authors
  1. Is the SASRec baseline reported in Table 2 exactly the same trained model checkpoint used as the Estimator for SASRec+SL?
  2. For the RHO-LOSS baseline, is the backbone trained on the full training set or only on the remaining subset after excluding the held-out data used to train the estimator? What is the held-out ratio? Is the backbone trained with the same masking ratio and training budget as SL4SR?
  3. What loss functions are used to train the estimator and the backbone? What negative sampling strategy is employed?
  4. The total training cost (including estimator pre-training and backbone training) should be reported. Besides, a comparison with SL4SR and RHO-LOSS under the same computational budgets would help clarify the practical efficiency of SL4SR.
2 comments

We sincerely appreciate the time and effort. We hope our responses address your concerns.

x\phantom{x}

W1 & Q2. Methodological Distinction and RHO-LOSS Implementation

The key distinction between SL4SR and RHO-LOSS lies in the role of the reference model and the motivation for its training strategy. RHO-LOSS employs a holdout-trained reference model to approximate irreducible loss, whereas SL4SR uses a pre-trained estimator to assess the current backbone's relative fitting behavior. In sequential recommendation, where user interaction data is often sparse, training the estimator on the full dataset allows it to leverage all available supervision and identify instances that the backbone has not yet sufficiently learned. Importantly, the estimator guides instance selection without directly transferring its predictions, while the backbone is optimized using ground-truth targets.

Regarding the RHO-LOSS implementation, we reserved 20% of the training instances as a disjoint holdout for reference model pre-training and used the remaining 80% for backbone optimization. Both methods employed the same backbone and estimator architectures, model configurations, optimizer, learning rate, batch size, maximum number of epochs, and early-stopping protocol. In Table 3, we applied the same masking ratio (ρ\rho) to both methods within each dataset–backbone combination, retaining the top (1-ρ\rho) fraction of candidate instances, corresponding to the RHO-LOSS selection fraction (nb/nBn_b/n_B).

The difference in training data usage reflects the respective methodological designs of the two approaches. RHO-LOSS relies on a disjoint holdout for reference loss estimation, whereas SL4SR intentionally uses the full training set to establish an estimator-relative fitting reference. Accordingly, our comparison preserves each method's training protocol while matching the model configurations and sample selection ratios. We clarify these settings to ensure transparency and reproducibility.

x\phantom{x}

W2. Comparison with Previously Reported Backbone Performance

Although the original FMLPRec paper also reports full-ranking results, absolute performance may vary across execution environments and experimental runs, even under similar experimental settings. We implemented the backbone models based on their publicly released code, preserving their original architectures and configurations.

Importantly, SL4SR is a model-agnostic training framework, and our primary objective is to evaluate the effectiveness of selective learning when applied to the same backbone under identical experimental conditions. To ensure fair comparisons, we use consistent model architectures, datasets, and evaluation protocols for each backbone and its SL-enhanced variant. Therefore, the performance improvements in Table 2 reflect the effect of the proposed selective learning strategy under controlled settings. The implementation details and experimental configurations are also available in our publicly released code for reproducibility.

x\phantom{x}

Q1. SASRec Baseline and Estimator Checkpoint

Yes. The SASRec estimator used for SASRec+SL is initialized from exactly the same trained checkpoint as the SASRec baseline reported in Table 2. The estimator remains frozen throughout selective learning, while the SL-enhanced backbone is trained separately using estimator-guided sample selection.

x\phantom{x}

Q3. Loss Functions and Negative Sampling Strategy

Both the estimator and backbone are trained using the cross-entropy loss for next-item prediction. We do not employ negative sampling and evaluate recommendation performance using full-item ranking over the entire item set.

x\phantom{x}

Q4. Training Cost and Computational Efficiency

Appendix B.5 (Figure 10) reports the per-epoch training runtime of the estimator and selective backbone. We further report the estimated total training time, including reference model pre-training and backbone optimization, alongside recommendation performance in Table C below. Both methods require relatively short training times with early stopping, making their computational costs practically manageable. Although SL4SR requires more training time in some settings, it consistently achieves higher HR@10 and NDCG@10 than RHO-LOSS and even requires less time on LastFM, demonstrating a favorable trade-off between training cost and recommendation accuracy.

Table C: Total Training Cost and Recommendation Performance. Ref./Back. denotes the number of training epochs for reference model pre-training and backbone optimization, respectively.

DatasetMethodEpochs (Ref./Back.)Total Time (s)HR@10/NDCG@10
BeautyRHO-LOSS59/76753.10.0492/0.0251
SL4SR42/100991.60.0567/0.0301
SportsRHO-LOSS11/65878.90.0287/0.0151
SL4SR51/1131705.70.0308/0.0168
LastFMRHO-LOSS13/68107.20.0624/0.0367
SL4SR23/5293.50.0670/0.0380

Dear eR5i,

We have provided responses addressing your concerns regarding (1) methodological novelty, (2) baseline performance, and (3) implementation details and computational efficiency. If you have any remaining concerns or questions, please let us know. We would be happy to provide further clarification.

Best regards,

Authors

Reviewer_zbPEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents SL4SR, a selective learning framework for sequential recommendation that aims to improve model generalization by selecting informative training instances. Inspired by prior work on sample selection in machine learning, SL4SR introduces a pre-trained estimator as a reference model and uses the difference between the backbone loss and the estimator loss to calculate an estimator-guided score for each training instance. During training, instances with higher scores are retained for updating the backbone, while lower-scoring instances are excluded. The main contribution of this work is the adaptation and evaluation of selective learning in sequential recommendation. Experiments on six real-world datasets and four representative backbones, including SASRec, GRU4Rec, FMLPRec, and Mamba4Rec, show consistent performance improvements. The ablation studies and analyses of the masking ratio further support the effectiveness of the proposed method.

Critical strengths
  1. The paper shows that selecting informative training instances can improve sequential recommendation performance. The results across multiple datasets, along with comparisons against random masking and RHO-LOSS, support the effectiveness of estimator-guided sample selection in recommendation area.

  2. The proposed framework can be applied to different sequential recommendation backbones without modifying their architectures. The performance improvements across SASRec, GRU4Rec, FMLPRec, and Mamba4Rec demonstrate its applicability to different model architectures. In addition, the framework introduces no additional inference overhead, which partly supports its practical value.

Critical weaknesses

1.Limited methodological novelty.

The core idea of SL4SR is highly similar to the existing RHO-LOSS method. Both approaches measure the learning utility of training instances by computing the sample-wise loss difference between the current model and a pre-trained reference model, and subsequently select high-scoring instances for parameter updates. Although applying this strategy to sequential recommendation has some practical value, SL4SR does not introduce a new sample scoring criterion or selection mechanism specifically tailored to the unique characteristics of sequential recommendation. Compared with RHO-LOSS, the main difference lies in how the reference model is trained, rather than in the core algorithmic design. Therefore, the contribution of this paper primarily consists of applying and empirically validating an existing method in a new task, with rather limited substantive methodological innovation.

2.Insufficient justification for the modifications to RHO-LOSS.

RHO-LOSS is based on the concept of reducible holdout loss and constructs reference losses using an independent holdout dataset or a cross-fitting strategy, thereby providing a basis for estimating the generalizable learning utility of training instances. In contrast, SL4SR directly pre-trains its estimator on the full training set and uses the estimator's losses on the same training instances to guide sample selection. Since the estimator has already been trained on the instances being evaluated, its low losses may partly result from memorizing specific training examples rather than capturing patterns that generalize to unseen data. The paper does not adequately explain why this modification is theoretically more justified. Given that this change represents the primary methodological difference between SL4SR and RHO-LOSS, the lack of sufficient justification further weakens the rationale behind the proposed method.

3.Concerns regarding the stability and reporting of experimental results.

Although the paper reports performance improvements across multiple datasets and backbones, some improvements are relatively small, and their stability remains unclear. Moreover, the optimal masking ratios vary considerably across different models and datasets. In some settings, the optimal ratio is as low as 0.01, meaning that approximately 99% of the training instances still contribute to optimization. This appears to be at odds with the paper's motivation of selectively filtering out less useful training instances. More importantly, the correspondence between the main results in Table 2, the optimal masking ratios in Table 5, and the sensitivity analyses in Figures 4 and 7 is unclear. For example, on the Sports dataset, Table 5 reports an optimal masking ratio of 0.01 for FMLPRec. However, according to the corresponding sensitivity curve, the HR@10 at this masking ratio does not appear to match the value of 0.0311 reported in Table 2. The authors should clearly specify the criteria used for hyperparameter and evaluation result selection, and confirm whether all test metrics for the same model were obtained under a single training configuration, to ensure the transparency and reproducibility of the experimental results.Very limited methodological novelty.

Questions for authors

1.The paper interprets a large loss gap between the backbone and the estimator as an indication of higher learning utility. However, this gap may also be influenced by differences in model fitting, biases in the reference model, and memorization of training instances. Could the authors further explain why this score can effectively reflect the generalization benefits that a training instance provides on unseen data?

2.RHO-LOSS constructs reference losses using an independent holdout dataset or a cross-fitting strategy, whereas SL4SR directly pre-trains its estimator on the full training set. Could the authors explain the theoretical motivation for this modification and provide controlled experiments that account for differences in the reference model's training data, model capacity, and degree of training, to clearly demonstrate the advantages of this design over RHO-LOSS? Furthermore, in Table 3, RHO-LOSS performs substantially worse than random masking or even the original backbone in several settings. Could the authors further explain this phenomenon and provide the implementation details and hyperparameter tuning procedures used for the RHO-LOSS baseline?

3.Could the authors clarify the criterion used to determine the optimal masking ratios in Table 5, and whether all metrics reported in Table 2 were obtained using the same masking ratio and checkpoint for each corresponding model? For example, the optimal masking ratio for FMLPRec on the Sports dataset is reported as 0.01, but the HR@10 near this setting in Figure 7 appears inconsistent with the value of 0.0311 reported in Table 2. Could the authors explain this discrepancy and provide the complete numerical results and corresponding validation metrics for different masking ratios?

4.SL4SR employs a pre-trained estimator to guide sample selection, but the estimator's fitting behavior may be influenced by item popularity and user interaction frequency. Could the authors analyze the item popularity distribution, user sequence length distribution, and selection consistency across different estimators for the selected training instances? In the cross-estimator experiments, using SASRec as the estimator for Mamba4Rec actually leads to a performance degradation. Does this suggest that the effectiveness of the loss gap is highly dependent on architectural compatibility between the estimator and the backbone, rather than reflecting the intrinsic informativeness of training instances?

5 comments

We sincerely thank you for your time and thoughtful comments. We hope our responses adequately address your concerns.

x\phantom{x}

W1. Methodological Novelty

We appreciate the reviewer's comments regarding the relationship between SL4SR and RHO-LOSS. While both methods employ loss discrepancies for sample selection, their underlying motivations and interpretations of the reference model differ.

RHO-LOSS primarily aims to improve training efficiency by estimating reducible holdout loss using the current model's loss and a reference model trained on independent data. In contrast, SL4SR addresses the uniform optimization paradigm in sequential recommendation (SR), where training instances may contain behavioral patterns with varying degrees of learnability and generalizability. Rather than estimating irreducible loss through a holdout model, SL4SR employs an estimator trained on the full training set as a reference for assessing how well the current backbone has learned each instance. The resulting loss discrepancy serves as a model-relative signal for prioritizing instances that may still provide useful learning opportunities.

Importantly, our contribution extends beyond applying an existing selection strategy to a different task. We establish estimator-guided selective learning as a model-agnostic optimization framework for SR and demonstrate its effectiveness across six datasets and four architecturally distinct backbones. These findings highlight the potential of revisiting training instance selection as a complementary direction to architectural improvements in SR.

x\phantom{x}

Q1. Interpretation and Effectiveness of the Loss-Gap Score

The loss-gap score is intended as a practical proxy for estimator-relative learnability, rather than a direct measure of intrinsic sample informativeness or guaranteed generalization benefit. Specifically, a large loss gap indicates that the current backbone fits an instance substantially less well than the pre-trained estimator, suggesting that the instance may contain learnable patterns that the backbone has not yet sufficiently captured. By comparing the two losses, SL4SR prioritizes instances based on their relative fitting discrepancy rather than their prediction difficulty alone.

We empirically examine the usefulness of this criterion through the score-tier analysis in Table 7. Under the same sample budget, selecting the Top 10% high-score instances substantially outperforms selecting the Middle or Bottom 10% for both SASRec and FMLPRec. Furthermore, the case study in Table 6 provides qualitative examples in which high-gap instances exhibit behavioral patterns consistent with subsequent user interactions. These findings support the effectiveness of the proposed score as a practical selection criterion for improving recommendation generalization.

W2 & Q2. Motivation for the Full-Data Estimator and Comparison with RHO-LOSS

Unlike RHO-LOSS, which uses a holdout-trained reference model to approximate irreducible loss, SL4SR employs a pre-trained estimator to assess the current backbone's relative fitting behavior. In sequential recommendation, where interaction data is often sparse, training the estimator on the full dataset allows it to leverage all available supervision and provide a reference for identifying instances that the backbone has not yet sufficiently learned.

Although memorization may influence the estimator's losses, the estimator does not directly transfer its predictions to the backbone. Instead, it guides instance selection, while the backbone is optimized using ground-truth targets. Thus, the loss discrepancy serves as a practical proxy for model-relative learnability, without assuming that it provides an unbiased estimate of generalization utility.

RHO-LOSS Implementation and Hyperparameter Evaluation. We implemented RHO-LOSS following its disjoint-holdout principle, using 20% of the training instances for estimator pretraining and the remaining 80% for backbone optimization. Both methods used the same backbone and estimator architectures, loss functions, and optimization settings. For a controlled initial comparison, we applied the same masking ratios to RHO-LOSS and SL4SR. To further ensure a comprehensive comparison, we independently tuned RHO-LOSS over the same masking-ratio search space as SL4SR (0.01–0.9), using the same validation metric, and report its best-performing configurations in Table A. Despite this tuning, SL4SR consistently outperformed RHO-LOSS across all eight comparisons, indicating that the performance gap is not solely attributable to the original masking-ratio settings.

Table A. Comparison with RHO-LOSS After Masking-Ratio Tuning

BackboneDatasetRHO RatioRHO(HR@10/NDCG@10)SL4SR(HR@10/NDCG@10)
SASRecBeauty0.10.0518/0.02650.0567/0.0301
Sports0.40.0287/0.01510.0308/0.0168
Instruments0.010.0994/0.05600.1155/0.0650
LastFM0.60.0624/0.03670.0670/0.0380
FMLPRecBeauty0.10.0508/0.02630.0592/0.0311
Sports0.30.0293/0.01520.0311/0.0161
Instruments0.010.1043/0.05330.1127/0.0600
LastFM0.40.0587/0.03420.0706/0.0413

Effect of Training-Data Allocation. To examine whether the reduced backbone training set accounts for the performance gap, we evaluated three RHO-LOSS allocations using SASRec: the original disjoint 20/80 split, a disjoint 50/50 split, and a 20/100 setting where the estimator uses 20% of the data while the backbone accesses the full training set, allowing overlap between the two. All configurations used the original matched masking ratios to isolate the effect of data allocation. As shown in Table B, increasing the disjoint holdout to 50% consistently reduced RHO-LOSS performance, while allowing full-data backbone training improved Beauty but not the other datasets consistently. SL4SR outperformed all three configurations, suggesting that the performance gap cannot be explained solely by backbone training-data availability. However, these comparisons do not isolate the independent effect of full-data estimator pretraining, and thus support the effectiveness of SL4SR's overall design rather than any single modification.

Table B. Effect of RHO-LOSS Training-Data Allocation Using SASRec

MethodBeautySportsInstrumentsLastFM
HR@10/NDCG@10HR@10/NDCG@10HR@10/NDCG@/10HR@10/NDCG@/10
RHO-LOSS (20/80)0.0492/0.02510.0287/0.01510.0784/0.04340.0624/0.0367
RHO-LOSS (50/50)0.0473/0.02390.0258/0.01340.0735/0.04040.0422/0.0236
RHO-LOSS (20/100)0.0534/0.02800.0285/0.01470.0812/0.04060.0615/0.0348
SL4SR0.0567/0.03010.0308/0.01680.1155/0.06500.0670/0.0380

W3 & Q3. Stability, Masking Ratio, and Experimental Consistency

The optimal masking ratio varies across datasets and backbones because the appropriate degree of selective learning depends on data characteristics and model-specific learning dynamics. Importantly, SL4SR does not aim to maximize the number of excluded instances, but to prioritize useful training signals while preserving sufficient supervision. Therefore, a small optimal masking ratio (e.g., 0.01) does not contradict our motivation, as even limited selective filtering may benefit optimization.

Regarding experimental consistency, HR@10 and NDCG@10 may reach their highest values at different masking ratios, as illustrated in Figure 7. To avoid ambiguity, we will revise Table 5 to consistently report the masking ratios corresponding to the highest NDCG@10 values in Figure 7. For each dataset–backbone configuration, all six test metrics in Table 2 were obtained using the same masking ratio and a single checkpoint. To further assess the statistical significance of the reported improvements, we will report p-values for the performance comparisons in Table 2.

x\phantom{x}

Q4. Selection Bias and Cross-Estimator Compatibility

To investigate potential selection bias, we analyzed the selected training instances on Beauty. The selection rates for head and tail items were 71.82% and 67.05%, respectively, while those for head and tail users were 72.10% and 67.87%. Although head groups exhibit slightly higher selection rates, substantial proportions of tail instances are also retained, consistent with the improvements observed across head and tail users in Figures 5 and 9. We further examined selection consistency for Mamba4Rec using Mamba4Rec and SASRec estimators. The resulting scores showed a Spearman correlation of 0.311, with a Top-60% overlap of 64.52%, compared with a random expectation of 59.77%. Since SL4SR relies on loss discrepancies for instance prioritization, differences in the loss characteristics and learning dynamics of the estimator and backbone may affect selection consistency. Although a unified BCE scoring loss is used in cross-estimator settings, architectural differences can still lead to different fitting behaviors, potentially explaining the performance degradation observed for Mamba4Rec in Table 4.

Reviewer_FFpfOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

Summary

The paper focuses on sequential recommendation. It argues that, at any point during training, some examples are more beneficial to learn from than others. To address this, the authors introduce a selective learning framework that uses a pre-trained estimator as a reference and relies on estimator-guided loss comparisons to approximate how learnable each sample is, enabling more informed instance selection during training.

Critical strengths

Strengths

  • The paper is straightforward and easy to follow.
  • Table 2 suggests performance improvements, though the authors should report whether the gains are statistically significant.
  • Experiments are run against well-known baselines.
  • The source code is available.
Critical weaknesses

Weaknesses

  • The contributions appear incremental rather than substantial. The motivation is weak.
  • Figure 1 is not a strong illustrative example and may not clearly support the paper’s core intuition.
  • The evaluation is limited to relatively small datasets, and it omits several more recent/strong baselines.
  • Multiple experimental components need clearer explanation or stronger justification. For instance, Figure 8 in Appendix B.3 does not communicate clear takeaways; the case study in Appendix B.6 might be better placed in the main paper (e.g., as part of the ablation/analysis), but the current explanations do not feel sufficiently supported.
Questions for authors

Questions

  • In the scenario depicted in Figure 1, wouldn’t the method risk repeatedly recommending the Romance option, reinforcing a narrow preference loop in the case of User 2? Similarly, why should we discard User 1?
  • Overall, I’m concerned about how meaningful the contribution is, whether the chosen baselines are strong/complete enough, and whether the ablation and analysis are sufficient to justify the claims.
3 comments

We sincerely appreciate your time and valuable feedback. We hope our responses clarify your concerns.

x\phantom{x}

W1. Contribution and Motivation of SL4SR

The contribution of our work goes beyond an incremental improvement. While prior sequential recommendation (SR) studies have primarily focused on improving sequence modeling architectures and representation learning, our work revisits SR from a different perspective: whether all training instances should contribute equally to model optimization. In practice, user interaction sequences often contain incidental or transient behaviors that may provide limited generalizable supervision, yet conventional training objectives typically treat all instances equally.

The central contribution of SL4SR lies in introducing an estimator-guided selective learning framework for SR, where the usefulness of each training instance is evaluated relative to the current backbone rather than determined solely by its prediction loss. Specifically, SL4SR employs a fixed estimator trained on the full training set and dynamically prioritizes instances exhibiting larger loss discrepancies, thereby focusing optimization on patterns that the current backbone has not yet sufficiently captured.

Although sample selection has been explored in broader machine learning contexts, SL4SR introduces a distinct training perspective for sequential recommendation by explicitly considering instance-wise learnability during optimization. Furthermore, the proposed framework can be applied to diverse SR architectures without modifying their internal structures or introducing additional inference overhead. We believe these contributions highlight the significance of selective learning as a complementary optimization strategy for SR.

x\phantom{x}

W2 & Q1. Motivation and Interpretation of Figure 1

Figure 1 is designed to illustrate that not all training instances are equally informative or generalizable, as user interaction sequences may contain both representative preference patterns and incidental or transient behaviors. Consequently, treating all instances equally during optimization may not be optimal. This observation motivates SL4SR, which introduces estimator-guided selective learning to identify instances that may provide greater learning benefit to the current model.

Importantly, Figure 1 is a conceptual illustration rather than a literal depiction of the proposed selection algorithm. The discarded and selected samples in Figure 1 are hypothetical examples intended to illustrate differences in sample informativeness, rather than actual selection decisions made by SL4SR. SL4SR does not select instances based on predefined item categories, such as Action or Romance, nor does it discard entire users. Instead, selection is dynamically determined by the loss discrepancy between the backbone and the pretrained estimator, as defined in Eq. (1). Therefore, even instances from the same user may be selected or masked differently throughout training, as illustrated in Figure 11.

Regarding the concern about repeatedly recommending Romance items, SL4SR operates during training and does not explicitly favor particular genres or enforce repetitive recommendations at inference time. The method prioritizes instances based on their estimator-relative learnability rather than the frequency or consistency of specific item categories. We will clarify Figure 1 to better distinguish its conceptual motivation from the actual selection mechanism.

x\phantom{x}

W3. Dataset Scale and Baseline Selection

We respectfully clarify that our experimental evaluation covers six widely used real-world datasets spanning multiple recommendation domains, including Amazon Beauty, Sports, Instruments, Yelp, LastFM, and MovieLens-1M. These datasets exhibit diverse characteristics in terms of interaction sparsity, user behavior, and sequence length. In particular, ML-1M contains approximately one million interactions, while LastFM and ML-1M provide relatively long interaction histories compared with the Amazon datasets.

More importantly, SL4SR is a model-agnostic training framework rather than a new recommendation architecture. Therefore, our primary evaluation objective is to examine whether selective learning consistently improves different backbone models under controlled comparisons.

To this end, we evaluate four representative backbones spanning distinct modeling paradigms: SASRec (self-attention), GRU4Rec (recurrent modeling), FMLPRec (frequency-domain filtering), and Mamba4Rec (state-space modeling). We believe these evaluations provide meaningful evidence for the effectiveness and architectural generality of SL4SR.

W4. Clarity of Experimental Analysis

We appreciate the reviewer’s suggestions regarding the experimental analysis. Figure 8 is intended to qualitatively illustrate how selective learning affects the learned item embedding space across different datasets.

Regarding the case study in Appendix B.6, Figure 11 illustrates how instance selection decisions change throughout training, highlighting the dynamic nature of SL4SR. We agree that this analysis could help readers better understand the proposed framework if presented in the main paper. We will relocate the case study to the main paper to strengthen the presentation of our experimental findings.

x\phantom{x}

Q2. Overall Contribution and Experimental Justification

We hope the above clarifications address the reviewer’s concerns. To summarize, the contribution of SL4SR is supported by three complementary aspects:

  1. Conceptual contribution: We revisit the conventional uniform training paradigm in SR and introduce estimator-guided selective optimization as an alternative learning strategy.
  2. General applicability: We demonstrate consistent improvements across six datasets and four architecturally distinct backbone models without additional inference overhead.
  3. Experimental validation: Our ablation studies compare selective learning with random masking, Top-1 selection, and RHO-LOSS (Table 3), while additional score-tier analysis (Table 7) provides further evidence that the proposed scoring criterion identifies useful training instances.

Taken together, these results support the effectiveness of estimator-guided selective learning beyond simply reducing the number of training samples.

Dear FFpf,

We have provided detailed responses to your concerns regarding (1) the contribution and motivation, (2) the illustration in Figure 1, and (3) the experimental evaluation. We hope our clarifications have addressed your concerns. If you have any remaining questions or concerns, we would be happy to provide further clarification.

Best regards,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.