Dual-Interest Sequential Product Recommendation With Multi-Granular SSM
cs.AI · cs.LG
Quick verdict
DSRec adds a time-gated SSM short-term branch to Mamba; gains are consistent but modest, with missing time-aware baselines and notable polish flaws.
Summary
The paper proposes DSRec, a sequential recommendation model that splits each item's representation into long-term and short-term interest embeddings to address item polysemy. The long-term branch augments item embeddings with a mean-pooled user-history aggregate and encodes the sequence with a Mamba block; the short-term branch uses an independently learned projection concatenated with quantile-binned inter-click time-interval embeddings, fed to a time-gated SSM whose state update interpolates between SSM output and the previous state. Detached residual cross-links exchange signals between branches before FFNs and an MLP prediction head trained with cross-entropy. Under leave-one-out evaluation on MovieLens-1M, Amazon-Beauty, and Amazon-Video-Games, DSRec leads seven baselines (Caser through SIGMA) on eight of nine metric-dataset cells — e.g., HR@10 0.3217 on ML-1M (+2.58% over SIGMA) and 0.1287 on Video-Games (+3.87%) — while Beauty MRR@10 sits 2.2% below Mamba4Rec. Ablations credit the dual embeddings, both encoders, and cross-fusion, and a Dual-Mamba control supports the asymmetric design. Inference takes 83 ms, between Mamba4Rec (72 ms) and SASRec (123 ms). Main limitations: the closest time-aware SSM baseline (Tim4Rec) is cited but not compared, significance testing is unexplained, the text contains internal inconsistencies, and no code is released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Items are polysemous in sequential recommendation, so encoding each item as separate long-term and short-term interest embeddings improves next-item prediction.
- An asymmetric dual-SSM (full-sequence Mamba plus a time-gated SSM) outperforms duplicating identical Mamba encoders, especially on sparse Amazon datasets.
- Detached residual cross-fusion between branches improves accuracy over independent branches (Table III).
- DSRec achieves state-of-the-art accuracy on three public benchmarks while keeping linear-time inference close to single-SSM baselines.
Strengths
Asymmetric design is tested against a proper control
The Dual-Mamba ablation (§IV-C.e, Table III) directly answers whether two identical encoders suffice: it loses HR@10 0.0871→0.0665 on Beauty and 0.1287→0.1093 on Video-Games while staying near-par on ML-1M (0.3200 vs 0.3217), supporting the claim that the time-gated SSM-S branch matters most on sparse data.
Honest negative metric
Beauty MRR@10 (0.0354) is reported below Mamba4Rec (0.0362) as -2.2% in Table II and left unstarred, rather than hidden — rare candor in SOTA-table papers.
Efficiency is measured, not just theorized
Practical tuning and robustness coverage
Hyperparameters are stated at implementation level
Weaknesses
Closest prior baselines missing
No statistical protocol behind significance stars
Narrative contradicts its own tables
Internal inconsistencies cluster
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This arXiv preprint (2609.21548, cs.AI/cs.LG; HKUST) proposes DSRec, a dual-branch sequential recommender. The authors make three claims: (1) items are polysemous — the same item plays different semantic roles in long-term preference versus short-term session contexts — so each item should carry two learned embeddings; (2) an asymmetric pair of state-space encoders (a full-sequence Mamba for the long-term branch, a time-gated SSM for the short-term branch) beats duplicating identical encoders; (3) residual cross-fusion yields state-of-the-art accuracy at near-SSM efficiency on MovieLens-1M, Amazon-Beauty, and Amazon-Video-Games. This is a method paper with benchmark evaluation; no code release is mentioned.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is DSRec?
DSRec is a sequential recommendation model that represents each item with two learned embeddings: a long-term interest embedding encoded by a Mamba state-space model, and a short-term interest embedding enriched with inter-click time intervals and encoded by a time-gated SSM. The two branches exchange information through residual cross-connections to predict the next item a user will interact with.
How well does DSRec perform compared to other models?
Under leave-one-out evaluation on MovieLens-1M, Amazon-Beauty, and Amazon-Video-Games, the paper reports DSRec best in 8 of 9 dataset-metric cells against seven baselines from Caser to SIGMA — for example HR@10 of 0.3217 on MovieLens-1M (+2.58% over SIGMA) and 0.1287 on Video-Games (+3.87%). One metric regresses: Beauty MRR@10 is 2.2% below Mamba4Rec.
Is DSRec faster than Transformer-based recommenders?
Per the paper's Table VI on Amazon-Beauty, DSRec inference takes 83 ms and 3.55 GB per epoch-batch — slower and heavier than Mamba4Rec (72 ms, 2.69 GB) but faster and lighter than SASRec (123 ms, 7.58 GB). Its theoretical complexity is linear in sequence length, like other SSM and RNN models, versus quadratic for Transformers.
Is code available for DSRec?
The paper describes a PyTorch implementation built on the RecBole framework with detailed hyperparameters (dimension 64, Adam lr 0.001, batch size 2048, SSM state dimension 32), but it does not provide any public code or data release link, so results currently cannot be verified from an official implementation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1, 3, 4, and 5 are described in text only; plotted values (length study, loss curves, robustness breakdowns) could not be verified from the extraction.
- The abstract is truncated mid-sentence in the source text ('DSRec outperf...'); conclusions were drawn from the body text.
- No appendix or supplementary material appears in the extraction; hyperparameter search ranges and seed policy are unknown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.