Research map / Architectures and efficiency / RNNs and normalization
Sparse Meets Dense: Unified Generative Recommendations with Cascaded Sparse-Dense Representations
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 COBRA, a generative recommender that cascades sparse semantic IDs and end-to-end-learned dense vectors in one Transformer decoder, with a BeamFusion inference rule combining beam scores and nearest-neighbor similarity.
Problem
Generative recommenders that predict only quantized semantic IDs lose fine-grained information and lag behind sequential dense retrieval in precision, while hybrid methods like LIGER use fixed, same-granularity dense representations.
Approach
Interleave sparse semantic IDs and dense vectors in the input sequence, predict the sparse ID first and then the dense vector conditioned on it, train the dense text encoder end-to-end with contrastive loss, and at inference fuse beam scores with cosine similarity (BeamFusion) to rank candidates.
Key results
- On Amazon Beauty, COBRA reports Recall@5 of 0.0537 vs. 0.0454 for TIGER (about 18% relative gain).
- On the industrial dataset, COBRA reaches Recall@500 of 0.3716, a 42.2% improvement over the w/o Dense variant.
- Online A/B test on Baidu advertising reports a 3.60% conversion increase and 4.15% ARPU increase.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| integrate cascaded sparse IDs and dense vectors | COBRA · Beauty | Recall@5: 0.0454 -> 0.0537 | Helped |
| integrate cascaded sparse IDs and dense vectors | COBRA · Sports and Outdoors | Recall@5: 0.0264 -> 0.0305 | Helped |
| integrate cascaded sparse IDs and dense vectors | COBRA · Toys and Games | Recall@5: 0.0521 -> 0.0619 | Helped |
| add sparse IDs (ablation) | COBRA · Baidu Industrial Dataset | Recall@800: 0.3111 -> 0.4466 | Helped |
| add dense vectors (ablation) | COBRA · Baidu Industrial Dataset | Recall@800: 0.3273 -> 0.4466 | Helped |
| add BeamFusion (ablation) | COBRA · Baidu Industrial Dataset | Recall@800: 0.2855 -> 0.4466 | Helped |
| increase the BeamFusion coefficient τ | COBRA · recall-diversity analysis | diversity | Hurt |
Limitations
- Industrial dataset and online results are proprietary, limiting reproducibility.
- Method description assumes a single-level sparse ID for brevity; multi-level extension is not fully evaluated on public datasets.
- BeamFusion coefficients (tau, psi) are tuned by hand on a fixed phi, with trade-offs shown only for one setting.
Experimental setup
Datasets: Amazon Product Reviews (Beauty, Sports and Outdoors, Toys and Games) Baidu Industrial Dataset
Benchmarks: Amazon Beauty Amazon Sports and Outdoors Amazon Toys and Games Baidu Industrial Dataset
Metrics: Recall@K NDCG@K diversity (number of distinct IDs)
Baselines: P5 Caser HGN GRU4Rec BERT4Rec FDSA SASRec S3-Rec TIGER RecFormer
Builds on: TIGER RQ-VAE LIGER SASRec RecFormer
Closest papers by meaning
- Learning to Rank in Generative Retrieval (AAAI 2024 · similarity 0.87)
- Generative Retrieval Meets Multi-Graded Relevance (NeurIPS 2024 · similarity 0.86)
- LIME: Link-based User-item Interaction Modeling with Decoupled XOR Attention for Efficient Test Time Scaling (NeurIPS 2026 · similarity 0.86)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.86)
- Rethinking Semantic ID Construction for Generative Recommendation: SimHash with Parallel Decoding and Semantic Alignment (NeurIPS 2026 · similarity 0.85)
- UxSID: Semantic-Aware User Interests Modeling for Ultra-Long Sequence (NeurIPS 2026 · 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.