acceptodds
Under review as a conference paper at ICLR 2027

MiRec: Multi-scale Interest Decomposition and Mixing for Sequential Recommendation

Abstract

Sequential recommendation predicts the next item a user will interact with by modeling the evolution of their interests from historical interactions. Multi-scale modeling captures user interests over different historical ranges but still faces two challenges. At the intra-scale level, stable and dynamic interests are entangled in holistic representations, making them difficult to model separately. At the cross-scale level, fusing entire representations does not adequately distinguish the mixing requirements of these two types of interests, limiting the targeted integration of complementary information. To address these challenges, we propose MiRec, a multi-scale interest decomposition and mixing model. Specifically, intra-scale interest decomposition extracts stable and dynamic interest representations through causal smoothing and residual computation. The cross-scale interest mixer then integrates representations of the same type from different scales through learnable subspace exchange. Finally, the updated representations are fused for next-item prediction. Experiments on four public datasets show that MiRec achieves relative improvements of 1.61%–5.81% in HR@10 and 2.72%–5.65% in NDCG@10 over the strongest baselines. Our code is available at https://anonymous.4open.science/r/MiRec-6B25/.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Sequential Recommendation, Multi-scale Interest Modeling, Interest Decomposition, Cross-scale Interaction, Subspace Exchange

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_Cp6bOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a multi-scale interest decomposition and mixing model for sequential recommendation, aiming to capture users’ stable and dynamic interests and their interactions. The model constructs multi-scale sequence representations using historical windows of different lengths, extracts stable and dynamic interests through smoothing and residual computation, and introduces a cross-scale interest mixer to facilitate information exchange across temporal scales. Finally, it integrates the cross-scale updates with the original representations and employs scale aggregation and a gating mechanism for next-item prediction. Comparative experiments on multiple datasets support the effectiveness of the proposed method.

Critical strengths
  1. The paper provides relatively comprehensive experimental validation. It compares the proposed method with 17 baselines on four public datasets and further conducts ablation studies, sensitivity analyses, visualization analyses, computational efficiency analyses, and statistical significance tests. These experiments validate the effectiveness of the proposed method from multiple perspectives, and the results also demonstrate a certain degree of practical value.

  2. The paper is generally well-written, with clear and detailed descriptions of the proposed methodology. The appendix provides relatively comprehensive experimental settings, implementation details, and supplementary experiments. In addition, the authors have made their code publicly available, making the work generally easy to understand and reproduce.

Critical weaknesses
  1. The paper lacks sufficient novelty. The overall technical approach is quite similar to TimeMixer (https://openreview.net/pdf?id=7oLshfEIC2 ICLR 2024) in the time series forecasting domain, including multi-scale modeling and cross-scale information mixing. Although transferring these ideas to sequential recommendation has some merit, the paper does not introduce sufficient substantive innovations beyond this adaptation. Moreover, the underlying research problem (personalized modeling) has already been extensively studied. In my view, the paper lacks sufficient novelty in both its research problem and technical methodology, falling well short of the level of methodological innovation expected at ICLR.

  2. The paper provides insufficient details regarding experimental fairness. The authors do not adequately specify the training configurations, hyperparameter search procedures, or model selection strategies for the baseline methods, raising concerns about the fairness of the comparisons. In addition, the publicly available code repository lacks the evaluation.py file, making it currently impossible to directly reproduce the complete experimental results. Given that the paper relies heavily on comparisons with numerous baselines to substantiate the effectiveness of the proposed method, these issues require clear explanations and clarification.

Questions for authors
  1. Could the authors clarify what substantive methodological innovations this paper introduces compared with TimeMixer? Beyond transferring ideas from the time series forecasting domain to sequential recommendation, could the authors further explain the core innovations of this work and its main differences from existing methods?

  2. Could the authors provide more details on the experimental configurations of the baseline methods to help assess the fairness of the comparisons?

  3. The publicly available code repository lacks the necessary evaluation files, making it currently impossible to reproduce the complete experimental results. Could the authors provide further clarification and the necessary instructions for reproduction?

8 comments
AuthorsOpenReview ↗

Response to Reviewer Cp6b - Q1 (1/2) Main innovations relative to TimeMixer

Sequential recommendation requires distinguishing interest components with different patterns of variation and integrating complementary information from different historical ranges. TimeMixer decomposes multiscale sequences into seasonal and trend components and mixes them in opposite directions. Inspired by this design, we introduce a decomposition-based multiscale mixing framework into sequential recommendation. Specifically, MiRec introduces two key innovations: causal multiscale modeling for position-wise prediction and Sinkhorn-based cross-scale feature subspace exchange. These address causal adaptation and subspace interaction in multiscale mixing, respectively, jointly supporting the integration of information from different historical ranges. We further introduce a TimeMixer-SR baseline to evaluate the overall design in terms of recommendation performance, active parameter count, and per-epoch training time.

First, regarding causal modeling, sequential recommendation requires each position to predict the next item using only its visible history. The original TimeMixer models fully observed historical blocks, and its centered smoothing and temporal-axis mixing without prefix constraints do not directly satisfy this requirement. Therefore, replacing only the input embeddings and prediction head is insufficient for task adaptation. MiRec employs aligned causal windows and prefix smoothing to integrate information from different historical ranges while preserving causality at each position.

Second, regarding subspace interaction, different subspaces of recommendation representations can capture different aspects of latent information, whereas temporal-axis mixing does not explicitly model cross-scale correspondences between these subspaces. To address this limitation, MiRec introduces Sinkhorn-based feature subspace exchange to learn feature reorganization between components of the same type across adjacent scales. An approximate doubly stochastic constraint balances information allocation from source subspaces to target subspaces, thereby incorporating subspace-level interaction into multiscale mixing.

To validate these designs, we construct TimeMixer-SR, a recommendation adaptation that preserves TimeMixer’s core decomposition and mixing mechanisms while incorporating item embeddings and the next-item prediction task. We compare the two methods in terms of recommendation performance and computational cost. As shown in the table below, compared with TimeMixer-SR, MiRec achieves relative improvements of 10.41% and 15.56% in HR@10 and NDCG@10 on Beauty, respectively, and 3.78% and 5.94% on ML-1M. Meanwhile, with similar active parameter counts, MiRec reduces per-epoch training time by 58.77% and 89.88% on the two datasets, respectively. These results support the effectiveness of MiRec’s overall design and demonstrate its advantages in both recommendation performance and per-epoch training efficiency.

DatasetMethodHR@10 ↑NDCG@10 ↑Active Parameters (M) ↓Training Time (s/epoch) ↓
BeautyTimeMixer-SR0.45350.29952.432118.3928
BeautyMiRec0.50070.34612.42807.5833
ML-1MTimeMixer-SR0.77840.53740.229523.1348
ML-1MMiRec0.80780.56930.22542.3414
AuthorsOpenReview ↗

Response to Reviewer Cp6b - Q1 (2/2) Research questions and technical approaches in relation to existing personalization methods

We thank the reviewer for their comments on related work. MiRec’s key distinction is to decompose stable and dynamic interests within the same historical range, then integrate same-type components across historical ranges through structured subspace exchange. We explain this distinction through method design and experimental results.

(1) Method design: combining within-scale interest decomposition with structured cross-scale interaction.

History windows determine temporal coverage, while stable and dynamic interests characterize how components vary within each range. The following table summarizes the differences from representative methods.

MethodMain focus and techniquesSmoothing/residual decomposition at each scaleCross-scale subspace exchange with approximately doubly stochastic weights
ComiRec-SA (Cen et al., 2020) [1]Extracts multiple interest vectors through self-attention and balances recommendation accuracy and diversity——
IDURL (Lin et al., 2025) [2]Quantifies interest drift, learns drift representations, and performs cross-user alignment——
HyperHawkes (Peintner et al., 2025) [3]Models recurring long-term intents through Hawkes processes and short-term interests through attention, combining them at the prediction-score level——
MIMAR-SRec (Fang & Shi, 2026) [4]Models multi-granularity intents through prototype alignment, cross-user contrastive learning, and adversarial augmentation——
MiRecSeparates stable/dynamic components within each scale and integrates same-type interests across scales✓✓

MiRec constructs the two interest representations through causal smoothing and residual decomposition. At the same prediction position, it learns source-to-target subspace correspondences between adjacent scales to update the stable and dynamic branches separately. These architectural attributes distinguish its design from the compared methods.

(2) Experimental results: additional comparisons with multiscale and decomposition methods.

Results for the additional baselines are shown below. Each cell reports HR@10 / NDCG@10.

MethodBeautyML-1M
MiRec0.5007 / 0.34610.8078 / 0.5693
Pyramid Mixer [5]0.4482 / 0.29950.7466 / 0.5040
ScaleRec [6]0.4942 / 0.34540.7984 / 0.5556
DWTRec [7]0.4696 / 0.32520.7675 / 0.5340
IDCLRec [8]0.4709 / 0.32520.7534 / 0.4963
WPGRec [9]0.4563 / 0.30080.7006 / 0.4374

MiRec achieves higher results on both metrics across both datasets. Compared with ScaleRec, the strongest additional baseline, MiRec improves HR@10 and NDCG@10 by 1.32% and 0.20% on Beauty and 1.18% and 2.47% on ML-1M, respectively, supporting the effectiveness of its overall modeling framework.

The method design is presented in Sections 3.3–3.4, and the additional baseline results are provided in Appendix C.6.2, Table 9.

[1] Cen Y, Zhang J, Zou X, et al. Controllable multi-interest framework for recommendation. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 2020: 2942-2951.

[2] Lin X, Pan W, Ming Z. Towards interest drift-driven user representation learning in sequential recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2025: 1541-1551.

[3] Peintner A, Mohammadi A R, Müller M, et al. Hypergraph-based temporal modelling of repeated intent for sequential recommendation. In Proceedings of the ACM on Web Conference 2025. 2025: 3809-3818.

[4] Fang Y, Shi H. Multi-granularity intent modeling with adversarial robustness for sequential recommendation. In Proceedings of the AAAI Conference on Artificial Intelligence. 2026, 40(17): 14702-14710.

[5] Gong Z, Fan Z, Lu H, et al. Pyramid Mixer: Multi-dimensional multi-period interest modeling for sequential recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2025: 4380-4384.

[6] Li H, Yang Y, Zeng J, et al. Multi-Scale transformers with dual attention and adaptive masking for sequential recommendation. Information Processing & Management, 2026, 63(1): 104318.

[7] Lu S, Ge M, Zhang J, et al. Filtering with time-frequency analysis: an adaptive and lightweight model for sequential recommender systems based on discrete wavelet transform. In Proceedings of the International Conference on Intelligent Computing. 2025: 161-173.

[8] Choi Y, Lim C. Intent-interest disentanglement and item-aware intent contrastive learning for sequential recommendation. arXiv preprint arXiv:2501.07096, 2025.

[9] Liu P, Ji Z, Yan G. WPGRec: Wavelet Packet Guided Graph Enhanced Sequential Recommendation. In Proceedings of the International ACM SIGIR Conference on Research and Development in Information Retrieval. 2026: 3964–3969.

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

This paper proposes MiRec, a multi-scale sequential recommendation model that aims to improve how user interests are represented across different temporal scales. The method is motivated by the observation that stable and dynamic interests may remain entangled within each scale and that existing approaches may not sufficiently model interactions across scales. MiRec therefore constructs multi-scale representations using causal windows, decomposes each scale into smoothed and residual components corresponding to stable and dynamic interests, and then mixes these components across adjacent scales in opposite directions. Stable representations are propagated from coarse to fine scales, while dynamic representations are propagated from fine to coarse scales. Cross-scale interaction is implemented through a learnable subspace exchange mechanism with Sinkhorn-normalized weights, and the updated multi-scale representations are finally fused with the base sequence representation for next-item prediction. The method is evaluated on four public datasets against 17 baselines, together with ablation, sensitivity, visualization, and efficiency analyses.

Critical strengths

1. The paper offers a clear and insightful modeling perspective on multi-scale sequential recommendation.

MiRec separates temporal scale from interest dynamics and explicitly models stable and dynamic interests within every temporal scale. This gives the model a richer representation of how user interests evolve over time and leads naturally to the proposed cross-scale interaction scheme. The directional propagation design is particularly appealing, with stable information flowing from coarse to fine scales and dynamic information flowing from fine to coarse scales. This creates a coherent modeling pipeline in which the main components are closely tied to the paper’s motivation. The ablation results further support the usefulness of the decomposition, cross-scale mixing, progressive propagation, and directional design.

2. The empirical evaluation is extensive and provides strong support for the proposed method.

MiRec is evaluated on four public datasets against 17 baselines covering general sequential recommendation, multi-interest modeling, and temporal or multi-scale recommendation. It achieves the best reported HR@10 and NDCG@10 results across all four datasets, with consistent gains over strong recent baselines. The paper also goes well beyond the main comparison by including detailed ablations, alternative decomposition and mixing designs, representation visualizations, hyperparameter sensitivity studies, and efficiency analysis. These experiments give a fairly complete picture of both the overall performance of MiRec and the contribution of its key design choices.

Critical weaknesses

1. The evaluation leaves some uncertainty about how much MiRec benefits from genuinely distinct temporal scales.

MiRec uses four temporal windows of {1,4,16,64}, while the maximum sequence length is 50. In addition, Beauty and Video have relatively short average histories of 12.33 and 9.26, respectively. For short-history users, some of the larger windows may therefore cover the same effective history, which makes it difficult to assess how much complementary information is actually available across the coarse scales. This matters because cross-scale interaction is an important part of the paper’s contribution. The current experiments clearly support the effectiveness of the overall model, but a sequence-length-stratified analysis or an effective-scale analysis would make the multi-scale explanation considerably more convincing.

2. The interpretation of the stable/dynamic decomposition could be validated more directly.

The stable component is constructed through temporal smoothing, while the dynamic component is obtained from the corresponding residual. As a result, the observed differences in temporal smoothness between the two branches are partly encouraged by the design itself. The additional representation analyses are consistent with the intended interpretation, and the ablations provide good evidence that the decomposition is useful for recommendation performance. However, the current evidence does not fully establish that the two branches specifically capture stable and dynamic user interests, rather than more generally separating slowly varying and rapidly varying representation components. This limits my confidence in the interpretation of the mechanism, although it does not undermine the empirical effectiveness of the proposed decomposition. A more independent validation of the stable/dynamic interpretation would strengthen the paper.

Questions for authors

1. How often do the configured temporal scales correspond to genuinely different behavioral histories in the current datasets?

The maximum sequence length is 50, whereas MiRec uses windows {1,4,16,64}, and the reported average sequence lengths are only 12.33 on Beauty and 9.26 on Video. Could the authors report the distribution of effective history lengths and provide either sequence-length-stratified results or scale-removal results (e.g., {1,4}, {1,4,16}, {1,4,16,64})? This would help determine whether the improvements indeed rely on complementary information from distinct temporal scales.

2. Can the authors provide stronger evidence that the two branches specifically capture stable and dynamic user interests beyond the temporal behavior induced by smoothing and residual construction?

Since the stable component is obtained through temporal smoothing and the dynamic component is defined as the corresponding residual, smoother variation in the former and stronger local variation in the latter are partly expected by construction. The case studies in Appendix C.7 provide useful qualitative evidence that the stable branch can retain persistent preferences while the dynamic branch follows temporary interest changes. Could the authors provide a more systematic, dataset-level evaluation of this behavior across users? Such evidence would help distinguish the claimed stable/dynamic interest decomposition from a useful low-/high-variation temporal decomposition.

**3. Could the authors clarify the sampled-negative evaluation protocol, including the number of negative candidates and the sampling strategy? ** Since Table 1 reports some relatively small improvements over the strongest baselines, I am also interested in whether the advantage persists under full-catalog ranking (at least on a representative dataset). This would help assess whether the reported gains are robust to the candidate-set construction rather than specific to sampled evaluation.

4. Could the authors clarify how the diagonal/off-diagonal structure of the learned exchange matrices in Figure 4 should be interpreted, given that the feature subspaces are defined by contiguous slices of latent dimensions?

The ordering and basis of latent feature dimensions are not inherently identifiable, so it is unclear whether diagonal mass reflects a robust cross-scale correspondence pattern or partly the chosen feature parameterization. Does the stable-vs-dynamic difference in diagonal mass persist across different feature partitions or alternative subspace assignments? Such a control would strengthen the interpretation of the learned exchange matrices.

5. Could the authors clarify how much the final prediction actually relies on the proposed multi-scale representation rather than the base sequence representation?

MiRec computes the final representation as ot=gtht+(1−gt)zˉto_t=g_t h_t+(1-g_t)\bar z_t, where hth_t is the base sequence representation and zˉt\bar z_t is the aggregated multi-scale representation after decomposition and cross-scale mixing. Since the learned gate gtg_t can in principle place most of the weight on hth_t, could the authors report the distribution of gtg_t across datasets, and ideally across different history-length groups? A fixed-gate or no-bypass control would also help determine whether the proposed decomposition and mixing pathway is substantially used by the final predictor rather than serving mainly as an auxiliary representation.]

6 comments
AuthorsOpenReview ↗

Response to Reviewer BoQy– Q1 How often do the configured temporal scales correspond to genuinely different behavioral histories in the current datasets?

We thank the reviewer for this suggestion. We have added actual history-coverage statistics and scale-removal experiments. The results show that short histories primarily support comparisons between recent behavior and the complete visible history, while longer histories allow further distinctions between intermediate and broader historical ranges. The details are as follows.

(1) Actual window coverage depends on the user’s visible history length.

We have confirmed that the maximum input sequence length is 64 and aligned the configuration descriptions in the manuscript and code. Given nn visible nonpadding interactions, the initial pooling coverage of window ww is min⁡(w,n)\min(w,n).

For example, with 8 interactions, windows {1,4,16,64} cover 1, 4, 8, and 8 interactions, respectively, yielding three distinct historical ranges. With 20 interactions, they cover 1, 4, 16, and 20 interactions, yielding four distinct ranges. Windows 16 and 64 therefore provide different historical coverage when the visible history exceeds 16 interactions.

(2) Coverage statistics clarify the roles of scales across datasets.

We examine all eligible test users, retaining at most the latest 64 visible interactions and excluding padding positions. Under the full configuration {1,4,16,64}, the following table groups users by the window subsets representing distinct historical coverages.

DatasetWindows {1,4}Windows {1,4,16}Windows {1,4,16,64}
Beauty0.00%86.98%13.02%
Video33.95%58.47%7.58%
Kindle0.00%34.91%65.09%
ML-1M0.00%0.00%100.00%

Beauty and Video primarily provide contrasts between recent behavior and the complete visible history. Most Kindle users and all ML-1M users have four distinct historical ranges, allowing further differentiation among recent, intermediate, and broader historical information.

(3) Scale-removal experiments support the overall benefits of the complete multiscale configuration.

We compare three window configurations on Beauty and ML-1M, which have substantially different history-length distributions. Results are shown in the following table, with each cell reporting HR@10 / NDCG@10. Full MiRec uses the multi-run means reported in Table 1, while the reduced configurations use means from supplementary experiments.

Window setBeautyML-1M
{1,4}0.4846 / 0.33710.7913 / 0.5555
{1,4,16}0.4910 / 0.34000.8008 / 0.5624
MiRec {1,4,16,64}0.5007 / 0.34610.8078 / 0.5693

Adding window 16 improves both metrics on both datasets, supporting the inclusion of more complete historical information alongside local behavior. The full configuration achieves the highest reported means, supporting the complete multiscale design. Together with the coverage statistics, these results indicate that short histories primarily benefit from integrating local behavior with the complete visible history, while longer histories additionally support complementary modeling across distinct historical ranges.

These statistics and experimental analyses have been added to Section 4.6 and Appendix C.13, Table 15.

AuthorsOpenReview ↗

Response to Reviewer BoQy - Q2 Validating that the two branches capture stable and dynamic user interests

Thank you for raising this important question. Providing direct evidence for the semantic interpretation of stable and dynamic interests beyond differences in temporal smoothness has also been a persistent challenge that we have sought to address. Your suggestions provided a clear path forward and were very helpful.

Following your advice, we have added external movie-genre readout experiments and a real-user case to examine branch semantics through two content targets: persistent preference background and recent genre changes. The results support the stable branch’s emphasis on historical background and the dynamic branch’s emphasis on recent changes.

(1) Experimental setup: examining the interest content retained by each branch using external labels.

With the trained MiRec frozen, we evaluate both branches on two readout targets:

  • Historical background: the genre distribution of the visible history, excluding the latest eight interactions.
  • Recent change: the genre distribution of the latest four interactions minus that of the preceding four interactions.

Each movie contributes equal mass to its annotated genres. Both branches perform both tasks using equal-capacity linear readout heads and the same validation-based hyperparameter selection procedure. Genre labels are used only for post-hoc analysis and do not participate in recommender training. The auxiliary training, validation, and test users are disjoint.

(2) Cross-user results: the branches exhibit different strengths in recovering interest content.

The following table reports means and standard deviations over three model training seeds. Higher Skill indicates better recovery of target information relative to a constant predictor based on the mean training target. Bold highlights the better result between the two post-exchange branches for each task.

RepresentationHistorical background Skill ↑Recent change Skill ↑
Raw smoothed component0.2322 ± 0.00230.0375 ± 0.0021
Raw residual component0.0012 ± 0.00010.2902 ± 0.0017
Post-exchange stable branch0.4589 ± 0.00560.0136 ± 0.0050
Post-exchange dynamic branch0.0009 ± 0.00030.2423 ± 0.0007

Among 1,208 auxiliary test users, after averaging each user’s errors across the three seeds, the stable branch achieves lower background-readout error for 73.10% of users, while the dynamic branch achieves lower change-readout error for 71.19%. Both directions also hold in the aggregate comparisons for each of the 18 movie genres under every seed.

Compared with the raw smoothed component, the post-exchange stable branch achieves 29.53% lower background-readout error. The post-exchange dynamic branch obtains a recent-change Skill of 0.2423, below the raw residual component’s 0.2902 and above the post-exchange stable branch’s 0.0136, demonstrating its relative advantage in recovering recent changes.

(3) Real-user case: connecting the branch roles to specific preference content.

Auxiliary test user 5906 predominantly watches Drama throughout the earlier 43 interactions, with Drama shares of 0.6746 and 0.6364 in the first 21 and subsequent 22 interactions, respectively. Recent interactions include horror movies such as Nosferatu, Rosemary’s Baby, and Disturbing Behavior: the Horror share increases from 0.0000 in the preceding four interactions to 0.5000 in the latest four, while the Drama share decreases from 0.5833 to 0.2500.

Readouts at the same prediction position are shown in the following table. Values are averaged over three seeds; bold indicates the branch closer to the observed target.

Semantic targetObserved behavioral targetStable-branch readoutDynamic-branch readout
Historical Drama preference background0.65500.41520.2276
Recent Drama change−0.3333+0.0018−0.3422
Recent Horror change+0.5000+0.0168+0.2446

The stable branch’s background readout ranks Drama highest, while the dynamic branch captures both the increase in Horror and the decrease in Drama. Its Drama-change readout of −0.3422 is close to the observed change of −0.3333. All three seeds preserve these directions, connecting the cross-user findings to persistent preferences and recent content shifts.

The case analysis have been added in the revised manuscript Appendix C.10.

AuthorsOpenReview ↗

Response to Reviewer BoQy - Q3 The evaluation protocol and additional comparison under full-catalog ranking.

We thank the reviewer for the detailed suggestions regarding the evaluation protocol and robustness to candidate-set construction. We have clarified the sampled evaluation protocol and completed an additional full-catalog ranking comparison on the representative ML-1M dataset.

(1) Data splits and candidate sampling rules. Methods use the same splits, candidate construction, exclusions, and metric implementation. Sampling controls evaluation cost across models, datasets, and repeated runs; the rules are as follows.

ItemSetting
SplitChronological leave-one-out; penultimate/final interactions for validation/test
Input historyTraining history for validation; validation interaction additionally included for testing
Candidates1 target entry + 100 uniform draws with replacement; no deduplication
Sampling exclusionsOnly training-history items and padding ID 0
UsersRandomly sample 10,000 IDs if the count exceeds 10,000, otherwise use all; then skip users without valid histories or targets

(2) Additional comparison under full-catalog ranking.

We compare MiRec, SASRec, MRFF, and InDiRec on ML-1M under the paper’s sampled-ranking protocol and an additional full-catalog protocol. Both protocols cover all eligible ML-1M test users. The average results are as follows; sampled-ranking values are taken from the paper.

MethodSampled HR@10Sampled NDCG@10Full-catalog HR@10Full-catalog NDCG@10
MiRec0.80780.56930.20860.0971
SASRec0.79390.55140.19530.0920
MRFF0.79500.55090.19930.0945
InDiRec0.79320.55420.19690.0935

Under sampled ranking, MiRec improves HR@10 over MRFF by 1.61% and NDCG@10 over InDiRec by 2.72%, relative to the strongest baseline for each metric. Under full-catalog ranking, MRFF is the strongest baseline on both metrics, and MiRec achieves relative improvements of 4.65% and 2.75%, respectively, calculated from unrounded means. MiRec ranks first under both protocols, showing that its performance advantage on ML-1M is retained across sampled and full-catalog candidate construction.

The sampled-ranking results and protocol are reported in Table 1 and Appendix C.2, respectively.

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

MiRec is an ID-based sequential recommendation model that constructs representations using causal windows of different lengths. At each scale, it separates a smoothed component and a residual into stable and dynamic branches. It then progressively exchanges feature subspaces across adjacent scales, using coarse-to-fine propagation for the stable branch and fine-to-coarse propagation for the dynamic branch, with Sinkhorn-normalized exchange weights. The resulting representations are fused for next-item prediction. The paper reports improvements over 17 baselines on four datasets, together with extensive ablation, sensitivity, efficiency, and representation analyses.

Critical strengths
  1. The paper reports consistent improvements over 17 baselines across four datasets, providing empirical support for the approach under the evaluated settings.

  2. The experimental analysis is extensive, covering component ablations, parameter-matched controls, alternative mixing operators, sensitivity, and efficiency. These analyses help assess the contribution of individual design choices.

  3. The model explicitly combines information from different historical ranges through causal, temporally aligned representations. Separately processing smoothed and residual components offers a useful perspective on modeling user interests in sequential recommendation.

Critical weaknesses
  1. The incremental contribution does not yet clearly justify the added complexity. TimeMixer [1] already combines multiscale representations, moving-average/residual decomposition, and opposite mixing directions. MiRec's causal windows and feature-subspace exchange are meaningful differences, but their relationship to TimeMixer needs clearer discussion. Despite the extensive ablations, the modest gains over alternative exchange operators, without variability across independent runs, leave the necessity of the full design uncertain.

  2. The code release currently prevents reproduction: main.py imports a missing evaluation module, so even main.py --help fails. The default embedding dimension is also 64, whereas Appendix C.2 specifies 50. These appear fixable, but the missing evaluation code and inconsistent configuration limit verification of the reported results.

[1] Shiyu Wang et al. TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting. ICLR, 2024.

Questions for authors
  1. Please clarify the relationship to TimeMixer, specifically what is new beyond multiscale decomposition and the two opposing mixing directions. Which existing results best isolate the benefit of the proposed feature-subspace exchange and other added components over a simpler, comparably tuned alternative? If available, variability across independent runs for the full model and its closest exchange-operator alternatives would materially affect my assessment of whether the added complexity is justified.

  2. Can you provide the missing evaluation module and the exact commands/configurations used for the reported tables, resolving the 50-versus-64 embedding-dimension discrepancy? Please also specify the evaluation candidate sampling and exclusion rules and whether the compared methods use the same protocol.

  3. Please clarify the intended comparison scope with respect to recent generative sequential recommendation methods. Are any directly comparable under the same data and candidate-ranking protocol, or are there specific setting differences that explain their exclusion? I am asking for justification of the comparison scope, not treating the absence of every recent method as a standalone reason for rejection.

7 comments
AuthorsOpenReview ↗

Response to Reviewer mvpt - Q1 (1/3) Main innovations relative to TimeMixer

Sequential recommendation requires distinguishing interest components with different patterns of variation and integrating complementary information from different historical ranges. TimeMixer decomposes multiscale sequences into seasonal and trend components and mixes them in opposite directions. Inspired by this design, we introduce a decomposition-based multiscale mixing framework into sequential recommendation. Specifically, MiRec introduces two key innovations: causal multiscale modeling for position-wise prediction and Sinkhorn-based cross-scale feature subspace exchange. These address causal adaptation and subspace interaction in multiscale mixing, respectively, jointly supporting the integration of information from different historical ranges. We further introduce a TimeMixer-SR baseline to evaluate the overall design in terms of recommendation performance, active parameter count, and per-epoch training time.

First, regarding causal modeling, sequential recommendation requires each position to predict the next item using only its visible history. The original TimeMixer models fully observed historical blocks, and its centered smoothing and temporal-axis mixing without prefix constraints do not directly satisfy this requirement. Therefore, replacing only the input embeddings and prediction head is insufficient for task adaptation. MiRec employs aligned causal windows and prefix smoothing to integrate information from different historical ranges while preserving causality at each position.

Second, regarding subspace interaction, different subspaces of recommendation representations can capture different aspects of latent information, whereas temporal-axis mixing does not explicitly model cross-scale correspondences between these subspaces. To address this limitation, MiRec introduces Sinkhorn-based feature subspace exchange to learn feature reorganization between components of the same type across adjacent scales. An approximate doubly stochastic constraint balances information allocation from source subspaces to target subspaces, thereby incorporating subspace-level interaction into multiscale mixing.

To validate these designs, we construct TimeMixer-SR, a recommendation adaptation that preserves TimeMixer’s core decomposition and mixing mechanisms while incorporating item embeddings and the next-item prediction task. We compare the two methods in terms of recommendation performance and computational cost. As shown in the table below, compared with TimeMixer-SR, MiRec achieves relative improvements of 10.41% and 15.56% in HR@10 and NDCG@10 on Beauty, respectively, and 3.78% and 5.94% on ML-1M. Meanwhile, with similar active parameter counts, MiRec reduces per-epoch training time by 58.77% and 89.88% on the two datasets, respectively. These results support the effectiveness of MiRec’s overall design and demonstrate its advantages in both recommendation performance and per-epoch training efficiency.

DatasetMethodHR@10 ↑NDCG@10 ↑Active Parameters (M) ↓Training Time (s/epoch) ↓
BeautyTimeMixer-SR0.45350.29952.432118.3928
BeautyMiRec0.50070.34612.42807.5833
ML-1MTimeMixer-SR0.77840.53740.229523.1348
ML-1MMiRec0.80780.56930.22542.3414
AuthorsOpenReview ↗

Response to Reviewer mvpt - Q1 (2/3) Benefits of Feature Subspace Exchange and Other Added Components

We thank the reviewer for their attention to the sources of improvement from the added components. The most direct evidence comes from controlled comparisons of the exchange mechanism, explicit decomposition, and progressive updates, with Row Softmax and DualProj additionally matching parameter counts. The changes, evaluation purposes, and capacity-control status are summarized in the following table.

ControlChange and factor examinedConfirmed parameter match with MiRec
Fixed Exch.Fix subspace correspondences to assess the value of learning them—
Row SoftmaxReplace normalization only to assess source-side column balancing✓
Attn-MixUse content-dependent subspace attention as an alternative mixing operator—
DualProjFeed the same undecomposed representation into two independent projections to assess explicit decomposition✓
Non-prog. Mix.Use original source representations at every step to assess progressive updates—

Here, ✓ indicates confirmed parameter matching, while — indicates that parameter matching has not yet been confirmed.

(1) Feature subspace exchange: examining correspondence learning, balancing constraints, and the exchange operator.

Fixed Exch. and Row Softmax test learnable correspondences and normalization constraints, respectively, while Attn-Mix provides an alternative mixing operator. Row Softmax retains the same-sized learnable exchange matrices, subspace partitions, residual injection, and remaining architecture, replacing only Sinkhorn normalization with row-wise softmax. It therefore provides the most direct parameter-matched comparison for identifying the contribution of balancing constraints. Sinkhorn iterations introduce no additional learnable parameters.

(2) Explicit decomposition: distinguishing its benefits from those of a two-branch architecture.

DualProj retains two independent projections, both mixing paths, and the same parameter count, but feeds the same undecomposed representation into both projections. MiRec’s improvement over this variant directly supports separating interest components with different variation patterns before mixing them across scales.

(3) Progressive updates: examining the value of updated source representations.

Non-prog. Mix. uses the original source representation at every adjacent-scale mixing step, whereas MiRec propagates progressively updated source representations. This comparison examines the value of passing already-integrated information to adjacent scales and progressively aggregating information from different historical ranges.

Representative results on Beauty and ML-1M are shown in the following table. Each cell reports HR@10 / NDCG@10.

MethodBeautyML-1M
Fixed Exch.0.4858 / 0.33640.8000 / 0.5599
Row Softmax0.4888 / 0.33960.8040 / 0.5638
Attn-Mix0.4882 / 0.34040.8013 / 0.5635
DualProj0.4844 / 0.33660.8003 / 0.5624
Non-prog. Mix.0.4862 / 0.33880.8007 / 0.5636
MiRec0.5007 / 0.34610.8078 / 0.5693

MiRec outperforms these variants on both metrics across all four datasets. These component-level comparisons identify the benefits of learnable exchange, balancing constraints, explicit decomposition, and progressive updates, complementing the TimeMixer-SR experiment that evaluates the overall adaptation.

These results are reported in Section 4.3, Table 2, and Appendix C.7, Tables 10–12.

AuthorsOpenReview ↗

Response to Reviewer mvpt - Q1 (3/3) Multi-seed means and standard deviations.

Below are means ± sample standard deviations over three training seeds, covering the full model, close exchange alternatives, and other component controls. Representative Beauty and ML-1M results are shown.

MethodBeauty HR@10Beauty NDCG@10ML-1M HR@10ML-1M NDCG@10
Fixed Exch.0.4858 ± 0.00590.3364 ± 0.00410.8000 ± 0.00550.5599 ± 0.0065
Row Softmax0.4888 ± 0.00360.3396 ± 0.00270.8040 ± 0.00150.5638 ± 0.0032
Attn-Mix0.4882 ± 0.00300.3404 ± 0.00160.8013 ± 0.00240.5635 ± 0.0051
DualProj0.4844 ± 0.00450.3366 ± 0.00380.8003 ± 0.00370.5624 ± 0.0029
Non-prog. Mix.0.4862 ± 0.00560.3388 ± 0.00460.8007 ± 0.00490.5636 ± 0.0043
MiRec0.5007 ± 0.00530.3461 ± 0.00370.8078 ± 0.00060.5693 ± 0.0009

Table 2 (p. 8) and Appendix C.7, Tables 10–12 (pp. 21–22), provide means and sample standard deviations for all four datasets. MiRec has higher means than these variants on both metrics across all four datasets, supporting the components, while standard deviations expose the corresponding variability.

Reviewer_6WXEOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes MiRec, a multi-scale interest decomposition and mixing model for sequential recommendation. It addresses two challenges in existing multi-scale methods: the entanglement of stable and dynamic interests within each temporal scale, and the limited interaction between interests across different scales.

MiRec first constructs representations from historical interactions using multiple causal windows. At each scale, it applies causal smoothing and residual computation to obtain separate stable and dynamic interest representations. It then introduces a cross-scale interest mixer that partitions these representations into feature subspaces and learns their correspondence through Sinkhorn-normalized exchange matrices. The two branches are updated progressively in opposite directions: stable interests propagate from coarse to fine scales, while dynamic interests propagate from fine to coarse scales. The resulting representations are fused with the base sequence representation for next-item prediction.

Experiments on four public datasets show that MiRec outperforms the strongest reported baselines, with relative improvements of 1.61%–5.81% in HR@10 and 2.72%–5.65% in NDCG@10. The paper also provides ablation studies and analyses of the proposed decomposition and cross-scale mixing mechanisms.

Critical strengths

(1) A coherent architecture for multi-scale interest modeling. MiRec goes beyond simply aggregating representations from different temporal windows. It combines intra-scale decomposition with cross-scale subspace exchange, and uses different propagation directions for the stable and dynamic branches. This provides a coherent modeling framework for integrating information across temporal scales and gives the proposed method a clear architectural rationale.

(2)Broad empirical evaluation with meaningful component-level analysis. The paper evaluates MiRec on four public datasets against 17 baselines and reports consistent improvements across HR@10 and NDCG@10. The ablation studies go beyond removing individual components and compare alternative decomposition and mixing strategies. These experiments provide useful evidence for the effectiveness of the overall design and make the empirical evaluation more informative than a simple performance comparison.

Critical weaknesses

(1) The algorithmic contribution appears incremental relative to existing multi-scale architectures. MiRec builds on several established ideas, including multi-scale representation learning, smoothing-residual decomposition, and bidirectional cross-scale information propagation. In particular, its overall design shares notable similarities with TimeMixer, which also decomposes representations and propagates different components across scales in opposite directions. MiRec's main distinction lies in its Sinkhorn-based feature subspace exchange. However, the paper does not sufficiently establish that this mechanism constitutes a substantial algorithmic advance over existing cross-scale mixing approaches, rather than a more constrained way of parameterizing feature interactions. This leaves the novelty and significance of the central contribution insufficiently clear.

(2)The proposed interest decomposition and subspace exchange are not sufficiently justified at the mechanistic level. Causal smoothing and residual computation produce components with different temporal characteristics, but this alone does not establish that they correspond to semantically distinct stable and dynamic interests. The model also lacks explicit constraints ensuring that the two branches capture complementary rather than overlapping information. While the ablation studies show that the proposed components improve recommendation performance, they do not fully validate the claimed semantic separation or explain why Sinkhorn-constrained subspace exchange is necessary compared with simpler mixing alternatives. This weakens the justification for the central modeling assumptions behind MiRec.

Questions for authors

(1) Relationship to existing multi-scale architectures. How does MiRec fundamentally differ from existing multi-scale decomposition and mixing architectures, particularly TimeMixer's PDM? Both approaches decompose representations and propagate different components across scales in opposite directions. Is the main methodological novelty of MiRec the Sinkhorn-based feature subspace exchange? If so, what distinctive modeling capability does this mechanism provide beyond replacing the mixing operator with a constrained learnable mapping, and which results in the paper best demonstrate this distinction?

(2)Validity of the proposed decomposition and exchange mechanisms. The paper interprets the causally smoothed component and its residual as stable and dynamic interests, respectively. What evidence supports this semantic interpretation, rather than merely showing that the two components capture different temporal variation patterns? Furthermore, how do the reported comparisons isolate the contribution of the Sinkhorn constraint itself from the additional flexibility introduced by the exchange mechanism? Are the compared mixing alternatives controlled for model capacity?

8 comments
AuthorsOpenReview ↗

Response to Reviewer 6WXE

Response to Reviewer 6WXE - Q1

Thank you for your suggestions. We address the four points below, covering main innovations relative to TimeMixer,the comparison with related architectures, innovations for sequential recommendation, the modeling capability of feature subspace exchange, and the supporting evidence in the paper.

Response to Reviewer 6WXE - Q1 (1/4) Main innovations relative to TimeMixer

Sequential recommendation requires distinguishing interest components with different patterns of variation and integrating complementary information from different historical ranges. TimeMixer [1] decomposes multiscale sequences into seasonal and trend components and mixes them in opposite directions. Inspired by this design, we introduce a decomposition-based multiscale mixing framework into sequential recommendation. Specifically, MiRec introduces two key innovations: causal multiscale modeling for position-wise prediction and Sinkhorn-based cross-scale feature subspace exchange. These address causal adaptation and subspace interaction in multiscale mixing, respectively, jointly supporting the integration of information from different historical ranges. We further introduce a TimeMixer-SR baseline to evaluate the overall design in terms of recommendation performance, active parameter count, and per-epoch training time.

First, regarding causal modeling, sequential recommendation requires each position to predict the next item using only its visible history. The original TimeMixer models fully observed historical blocks, and its centered smoothing and temporal-axis mixing without prefix constraints do not directly satisfy this requirement. Therefore, replacing only the input embeddings and prediction head is insufficient for task adaptation. MiRec employs aligned causal windows and prefix smoothing to integrate information from different historical ranges while preserving causality at each position.

Second, regarding subspace interaction, different subspaces of recommendation representations can capture different aspects of latent information, whereas temporal-axis mixing does not explicitly model cross-scale correspondences between these subspaces. To address this limitation, MiRec introduces Sinkhorn-based feature subspace exchange to learn feature reorganization between components of the same type across adjacent scales. An approximate doubly stochastic constraint balances information allocation from source subspaces to target subspaces, thereby incorporating subspace-level interaction into multiscale mixing.

To validate these designs, we construct TimeMixer-SR, a recommendation adaptation that preserves TimeMixer’s core decomposition and mixing mechanisms while incorporating item embeddings and the next-item prediction task. We compare the two methods in terms of recommendation performance and computational cost. As shown in the table below, compared with TimeMixer-SR, MiRec achieves relative improvements of 10.41% and 15.56% in HR@10 and NDCG@10 on Beauty, respectively, and 3.78% and 5.94% on ML-1M. Meanwhile, with similar active parameter counts, MiRec reduces per-epoch training time by 58.77% and 89.88% on the two datasets, respectively. These results support the effectiveness of MiRec’s overall design and demonstrate its advantages in both recommendation performance and per-epoch training efficiency.

DatasetMethodHR@10 ↑NDCG@10 ↑Active Parameters (M) ↓Training Time (s/epoch) ↓
BeautyTimeMixer-SR0.45350.29952.432118.3928
BeautyMiRec0.50070.34612.42807.5833
ML-1MTimeMixer-SR0.77840.53740.229523.1348
ML-1MMiRec0.80780.56930.22542.3414

[1] Wang S, Wu H, Shi X, et al. Timemixer: Decomposable multiscale mixing for time series forecasting. In Proceedings of the International conference on learning representations. 2024, 2024: 38626-38652.

Response to Reviewer 6WXE - Q1 (2/4) Relation to existing multi-scale decomposition and mixing architectures

To clarify the relationship between MiRec and existing multi-scale decomposition and mixing architectures, we have conducted a supplementary review of the relevant research and compared these approaches across six dimensions: whether they are designed for sequence recommendation, whether they employ explicit decomposition, the use of a temporal MLP, opposite cross-scale mixing, causal window alignment, and subspace exchange, as shown in the table below.

MethodSequential recommendationExplicit decompositionTemporal MLPOpposite cross-scale mixingAligned causal windowsSubspace exchange
Autoformer [2]—✓————
FEDformer [3]—✓————
DLinear [4]—✓————
DWTRec [5]✓✓————
TSMixer [6]——✓———
Pyramid Mixer [7]✓—✓———
ScaleRec [8]✓—————
TimeMixer [1] (original PDM)—✓✓✓——
MiRec✓✓—✓✓✓

As shown in the table, existing methods cover various aspects such as sequence decomposition, temporal dimension mixing and cross-scale interactions. Among these, DWTRec has applied explicit decomposition to sequence recommendation, TimeMixer combines explicit decomposition, a temporal dimension MLP and reverse cross-scale mixing, but its original architecture was not designed for sequence recommendation, nor does it incorporate the causal window alignment and subspace exchange mechanisms listed in the table.

We further conducted experimental comparisons between MiRec and Pyramid Mixer [7], ScaleRec [8], DWTRec [5], IDCLRec [9] and WPGRec [10] on the Beauty and ML-1M datasets, using the same datasets and evaluation metrics as for MiRec. As shown in the table below, MiRec achieved higher HR@10 and NDCG@10 scores than each of the new baselines on both datasets. On the Beauty dataset, MiRec achieved relative improvements of 1.32% and 0.20% in the two metrics compared to ScaleRec; on the ML-1M dataset, MiRec achieved relative improvements of 1.18% and 2.47% in the two metrics. These results support the effectiveness of MiRec’s holistic modelling approach relative to the multi-scale modelling and decomposition methods evaluated.

MethodBeauty HR@10 ↑Beauty NDCG@10 ↑ML-1M HR@10 ↑ML-1M NDCG@10 ↑
MiRec0.50070.34610.80780.5693
Pyramid Mixer0.44820.29950.74660.5040
ScaleRec0.49420.34540.79840.5556
DWTRec0.46960.32520.76750.5340
IDCLRec0.47090.32520.75340.4963
WPGRec0.45630.30080.70060.4374

We have expanded the discussion of existing architectures and additional baselines in Appendix C.6.1–C.6.2 of the manuscript (pp. 19–21, Tables 8–9). The relevant explanations and captions are marked in blue.

[2] Wu H, Xu J, Wang J, et al. Autoformer: Decomposition transformers with auto-correlation for long-term series forecasting. Advances in neural information processing systems, 2021, 34: 22419-22430.

[3] Zhou T, Ma Z, Wen Q, et al. Fedformer: Frequency enhanced decomposed transformer for long-term series forecasting. In Proceedings of the International conference on machine learning. PMLR, 2022: 27268-27286.

[4] Zeng A, Chen M, Zhang L, et al. Are transformers effective for time series forecasting? In Proceedings of the AAAI conference on artificial intelligence. 2023, 37(9): 11121-11128.

[5] Lu S, Ge M, Zhang J, et al. Filtering with time-frequency analysis: an adaptive and lightweight model for sequential recommender systems based on discrete wavelet transform. In Proceedings of the International Conference on Intelligent Computing. 2025: 161-173.

[6] Chen S A, Li C L, Yoder N, et al. Tsmixer: An all-mlp architecture for time series forecasting. Transactions on Machine Learning Research, 2023.

[7] Gong Z, Fan Z, Lu H, et al. Pyramid Mixer: Multi-dimensional multi-period interest modeling for sequential recommendation. In Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 2025: 4380-4384.

[8] Li H, Yang Y, Zeng J, et al. Multi-Scale transformers with dual attention and adaptive masking for sequential recommendation. Information Processing & Management, 2026, 63(1): 104318.

[9] Choi Y, Lim C. Intent-interest disentanglement and item-aware intent contrastive learning for sequential recommendation. arXiv preprint arXiv:2501.07096, 2025.

[10] Liu P, Ji Z, Yan G. WPGRec: Wavelet Packet Guided Graph Enhanced Sequential Recommendation. In Proceedings of the International ACM SIGIR Conference on Research and Development in Information Retrieval. 2026: 3964–3969.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.