acceptodds
Under review as a conference paper at ICLR 2027

MARS: Multi-Rate Aggregation of Recency Signals for Sequential Recommendation

Abstract

Sequential recommendation requires both encoding the content of a user's history and deciding how that content should contribute to the next prediction. We study temporal readout as a modelling choice that can be controlled separately from sequence encoding. We introduce MARS (Multi-rate Aggregation of Recency Signals), a post-encoder operator that uses observed timestamps to construct K recency-weighted summaries. User-conditioned effective decay rates determine the temporal concentration of each head, and a context-dependent gate fuses the summaries with the encoder's final state. The same readout interface supports Transformer and Mamba backbones without changing downstream item scoring. Our analysis characterises a restricted kernel-approximation family and establishes how effective rates distinguish attention profiles and control their expected lag. In the reported results on four sparse benchmarks, the Transformer instantiation improves HR@10 by an average of 19.7% over the strongest content-only Transformer baseline on each dataset. On dense ML-1M, the Mamba instantiation exceeds SIGMA's HR@10 by 3.2%. Matched-backbone ablations show that gains depend on the encoder and dataset, while single-head variants remain competitive in several settings. These findings support explicit temporal readout as a modular means of introducing recency bias, with its value determined jointly by the encoded representation and the interaction history. Code is available at https://anonymous.4open.science/r/MARSS.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Sequential Recommendation, Temporal Readout, Multi-Rate Aggregation, Recency Modeling, Adaptive Fusion

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_wStsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this paper, the authors design a post-encoder aggregation module, MARS (Multi-rate Aggregation of Recency Signals), and integrate it into existing networks such as Transformer and Mamba. The authors conclude that a separate controlled modelling of temporal readout can be valuable for the sequential modeling task. Comprehensive experiments compare the proposed methods with multiple baseline models with varying architectures and show that the proposed methods achieve state-of-the-art performance across recommendation datasets while maintaining computational efficiency.

Critical strengths
  1. A major highlight of this work is the conceptual decoupling of temporal modeling from sequence representation. By designing MARS as a post-encoder aggregation operator rather than an in-encoder modification, the authors propose a highly modular solution. This decoupling not only yields a clean architectural design but also ensures that the proposed module can easily adapt to future advancements in underlying backbone networks.

  2. The proposed post-encoder module, MARS, can be applied to existing sequential modeling networks, and ablation experiments verify the performance gain caused by the proposed MARS. Importantly, the authors transparently report their experimental results, including instances where prior method, e.g., TiSASRec, maintain the strongest performance on specific datasets like the Games dataset. The authors declare that their experiments are from different rounds of running.

Critical weaknesses

Sequential modeling is not the reviewer's main research focus. The reviewer's main concerns are listed as follows.

  1. The efficiency comparisons inTable 5 show a counter-intuitive phenomenon: on datasets like Beauty and ML-1M, MARS-T achieves lower MFLOPs compared to the vanilla SASRec. Given that MARS-T is essentially SASRec appended with the proposed MARS module, the computational cost should theoretically be strictly greater than (or equal to) the base encoder. Could the authors clarify the underlying reason for this reduction in MFLOPs?

  2. The authors set the τ as a fixed constant number, 1 hour. Could the authors provide any insight about this parameter selection?

Questions for authors

Please see the weaknesses.

Reviewer_RbSLOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces MARS, a modular, post-encoder temporal readout operator for sequential recommendation. Instead of embedding temporal dynamics directly into the sequence encoder, MARS decouples temporal aggregation from content encoding. It constructs KK recency-weighted summaries using log-compressed elapsed times, where each decay head's effective rate is modulated by the user's encoded history. These summaries are then fused via a context-adaptive gate, combined with a residual connection to the encoder's final state.

Critical strengths
  • The core strength of MARS is its decoupled design. By formulating temporal readout as a post-encoder operator, it provides a unified interface that can be seamlessly plugged into different backbone architectures without altering the downstream item-scoring mechanism.
  • It solves the problem of deep binding between time modeling and sequence encoders in existing methods, and provides a modular post-processing interface that allows time aggregation to be evaluated and replaced independently of encoders.
  • The inclusion of JSD-based diversity loss and batch-level load-balance penalties shows a deep understanding of the optimization dynamics of multi-head gating mechanisms.
Critical weaknesses
  • Many existing methods focus on the multiple time scales in the recommender systems. This paper lacks the related discussion about the independent time reader and their unique superiority than existing time-aware methods. To be specific, what are the defects of coupling the time reading with encoder dynamics? The readers may be struggle to understand this question. In addition, more time-aware recommendation baselines should also be added to the experiment to further show the method's effectiveness.
  • The threshold for backbone selection should be discussed in detail. The Beauty, Sports and Games are the datasets with short average sequence lengths. This raises a concern about the adaptivity of MARS facing different interaction lengths. In addition, the improvement seems slight in some datasets.
  • In Table 3, the KK values of NDCG@KK and MRR@KK are missing.
Questions for authors

what are the defects of coupling the time reading with encoder dynamics? The threshold for backbone selection are encouraged to be discussed in detail.

Reviewer_zTgEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method that adaptively learns user-specific temporal importance by assigning different weights to historical interactions based on their elapsed times. The method aims to address two issues: (1) actual time intervals between interactions may not be sufficiently reflected in sequential recommendation models, and (2) the importance of historical interactions may vary across users. A key feature of the proposed approach is that it introduces an additional post-encoder module without substantially modifying the underlying Transformer or Mamba architecture.

Critical strengths

S1. Unlike many existing approaches that incorporate temporal information directly into the encoder, the proposed method applies temporal weighting to the encoder outputs. This design allows the same module to be applied to both Transformer and Mamba backbones. S2. By employing multiple decay rates, the method can capture both short-term and long-term user behaviors. Furthermore, the user-conditioned decay mechanism allows the model to adaptively adjust decay rates for individual users.

Critical weaknesses

W1. Although the proposed methodology is conceptually reasonable, my main concern is the lack of consistent empirical support for its effectiveness. In some cases, the single-head variant outperforms the multi-head variant, and the benefits of user-conditioned decay rates are not consistent across datasets. Similarly, incorporating actual timestamps improves performance on some datasets but not on others. These inconsistent results raise concerns about the empirical robustness of the proposed components.

Questions for authors

Q1. The backbone selection criterion appears to be empirically determined based on the average sequence length. Do the authors expect this criterion to remain valid for datasets with intermediate average sequence lengths (approximately 20–100)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.