Research map / Architectures and efficiency / Attention and transformers
SIGMA: Selective Gated Mamba for Sequential Recommendation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SIGMA, a Mamba-based sequential recommender with a partially flipped bidirectional Mamba, a dense selective gate, and a GRU branch for short-term patterns, reporting gains over transformer, RNN, and SSM baselines on five datasets.
Problem
Transformer-based sequential recommenders have quadratic cost, while directly applying unidirectional Mamba to sequential recommendation loses future context and performs poorly on short (long-tail) user sequences.
Approach
Partially flip the first n items of the sequence (keeping the last r items in place) so a second Mamba pass gives bidirectional context while preserving recent short-term patterns; fuse the two directions with an input-dependent Dense Selective Gate and add a Feature Extract GRU branch for short sequences.
Key results
- SIGMA achieves HR@10 of 0.0629 on Yelp, 8.82% over the best baseline LinRec (0.0579), and 0.0986 on Beauty.
- Ablation on Beauty: removing FE-GRU drops HR@10 from 0.0986 to 0.0750, the largest drop among the components.
- Single-layer SIGMA runs at 68ms inference per batch on Beauty versus 123ms for SASRec, with GPU memory 2.89G vs 7.58G.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add PF-Mamba with partial flipping (ablation) | SIGMA · Beauty | HR@10: 0.0953 -> 0.0986 | Helped |
| Add the Dense Selective Gate (ablation) | SIGMA · Beauty | HR@10: 0.0954 -> 0.0986 | Helped |
| Add FE-GRU (ablation) | SIGMA · Beauty | HR@10: 0.0750 -> 0.0986 | Helped |
| Add PF-Mamba with partial flipping (ablation) | SIGMA · Sports | HR@10: 0.0711 -> 0.0735 | Helped |
| Add the Dense Selective Gate (ablation) | SIGMA · Sports | HR@10: 0.0713 -> 0.0735 | Helped |
| Add FE-GRU (ablation) | SIGMA · Sports | HR@10: 0.0622 -> 0.0735 | Helped |
| Add PF-Mamba with partial flipping (ablation) | SIGMA · ML-1M | HR@10: 0.3289 -> 0.3308 | Helped |
| Add the Dense Selective Gate (ablation) | SIGMA · ML-1M | NDCG@10: 0.1866 -> 0.1906 | Helped |
| Add FE-GRU (ablation) | SIGMA · ML-1M | HR@10: 0.3304 -> 0.3308 | Helped |
| Increase the number of SIGMA layers from 1 to 2 (ablation) | SIGMA · Beauty | HR@10: 0.0986 -> 0.0994 | Helped |
| Increase the number of SIGMA layers from 2 to 4 (ablation) | SIGMA · Beauty | HR@10: 0.0994 -> 0.0963 | Hurt |
| Use SIGMA instead of the strongest baseline | Sequential recommendation · Sports | HR@10: 0.0746 -> 0.0735 | Hurt |
Surprising findings
- Increasing SIGMA from 1 to 4 layers lowers HR@10 on Beauty (0.0986 to 0.0963) while inference cost rises sharply, which the authors attribute to overfitting.
- On Sports, SIGMA's HR@10 (0.0735) is lower than FEARec's (0.0746), so it is not the best on that metric.
Limitations
- Gains over the strongest baselines are small on some datasets (e.g., Games HR@10 +0.68%, ML-1M +0.76%), and SIGMA trails FEARec on Sports HR@10.
- Multi-layer stacking does not improve accuracy, leaving the depth/overfitting trade-off unresolved.
- Partial-flip hyperparameter r is set per dataset and the ablation and tuning are reported only on Beauty in the main text.
Experimental setup
Datasets: Yelp Sports Beauty ML-1M Games
Benchmarks: Yelp Sports Beauty ML-1M Games
Metrics: HR@10 NDCG@10 MRR@10 inference time GPU memory
Baselines: GRU4Rec BERT4Rec SASRec LinRec FEARec Mamba4Rec ECHO-Mamba4Rec
Builds on: Mamba Mamba4Rec SASRec RecBole Vision Mamba
Closest papers by meaning
- SLMRec: Distilling Large Language Models into Small for Sequential Recommendation (ICLR 2025 · similarity 0.87)
- LLM4RSR: Large Language Models as Data Correctors for Robust Sequential Recommendation (AAAI 2025 · similarity 0.86)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
- Customizing Language Models with Instance-wise LoRA for Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
- Enhancing Sequential Recommendation with Global Diffusion (AAAI 2025 · similarity 0.86)
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.