Research map / Large language models / Retrieval and knowledge
SRA-CL: Semantic Retrieval Augmented Contrastive Learning 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 SRA-CL, which uses DeepSeek-V3 semantic embeddings to retrieve positive user/item candidates for inter-user and intra-user contrastive learning, plus a learnable sample synthesizer, as a plug-and-play add-on to sequential recommenders.
Problem
Contrastive learning for sequential recommendation builds unreliable positive pairs: random augmentations corrupt user preference semantics, collaborative-signal clustering is unreliable under sparsity, and predefined pairing rules limit learning.
Approach
Use LLM-generated preference summaries (for users) and item summaries (for items), embedded with SimCSE-RoBERTa, to retrieve semantically similar users and items as candidate pools; a learnable attention-based synthesizer combines retrieved users into inter-user positives, while item substitution within semantic neighborhoods yields intra-user positive views.
Key results
- SRA-CL achieves HR@20 of 0.1282 on Yelp, 0.0823 on Sports, 0.1314 on Beauty and 0.1702 on Office, improving over the best baseline by up to 11.82% in HR@20.
- Plugging SRA-CL into GRU4Rec, SASRec and DuoRec raises HR@20 by 8.3% to 27.3% and NDCG@20 by 9.7% to 25.5%.
- Removing the semantic retrieval (w/o semantic) or LLM processing (w/o LLM) drops HR@20 on Yelp from 0.1282 to 0.1187 and 0.1190 respectively.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add inter-user contrastive learning (ablation) | SRA-CL · Yelp, Sports, Beauty, Office | HR@20: 0.1203 -> 0.1282 (Yelp); 0.0780 -> 0.0823 (Sports); 0.1273 -> 0.1314 (Beauty); 0.1621 -> 0.1702 (Office) | Helped |
| add intra-user contrastive learning (ablation) | SRA-CL · Yelp, Sports, Beauty, Office | HR@20: 0.1228 -> 0.1282 (Yelp); 0.0795 -> 0.0823 (Sports); 0.1279 -> 0.1314 (Beauty); 0.1619 -> 0.1702 (Office) | Helped |
| add both inter-user and intra-user contrastive objectives (ablation) | SRA-CL · Yelp, Sports, Beauty, Office | HR@20: 0.1101 -> 0.1282 (Yelp); 0.0745 -> 0.0823 (Sports); 0.1206 -> 0.1314 (Beauty); 0.1476 -> 0.1702 (Office) | Helped |
| add learnable inter-user sample synthesis (ablation) | SRA-CL · Yelp, Sports, Beauty, Office | HR@20: 0.1253 -> 0.1282 (Yelp); 0.0792 -> 0.0823 (Sports); 0.1273 -> 0.1314 (Beauty); 0.1617 -> 0.1702 (Office) | Helped |
| use semantic information for retrieval (ablation) | SRA-CL · Yelp, Sports, Beauty, Office | HR@20: 0.1187 -> 0.1282 (Yelp); 0.0772 -> 0.0823 (Sports); 0.1265 -> 0.1314 (Beauty); 0.1624 -> 0.1702 (Office) | Helped |
| use LLM-based text processing (ablation) | SRA-CL · Yelp, Sports, Beauty, Office | HR@20: 0.1190 -> 0.1282 (Yelp); 0.0781 -> 0.0823 (Sports); 0.1259 -> 0.1314 (Beauty); 0.1643 -> 0.1702 (Office) | Helped |
| replace direct semantic item substitution with learnable synthesis (ablation) | SRA-CL intra-user contrastive learning · Yelp, Sports, Beauty, Office | HR@20: 0.1282 vs 0.1276 (Yelp); 0.0823 vs 0.0825 (Sports); 0.1314 vs 0.1309 (Beauty); 0.1702 vs 0.1706 (Office) | No clear effect |
| add SRA-CL contrastive losses to the backbone | GRU4Rec, SASRec, DuoRec · multiple backbones | HR@20: 8.3% to 27.3% | Helped |
Surprising findings
- A learnable synthesizer gave no measurable gain for intra-user item substitution (learnable vs unlearnable nearly identical), so the learnable component is only used for inter-user pairs.
- Sampling 'hard' same-next-item positives essentially duplicates the next-item objective and adds no information, according to the authors' argument.
Experimental setup
Datasets: Yelp Amazon Sports Amazon Beauty Amazon Office
Benchmarks: Yelp Sports Beauty Office
Metrics: HR@20 NDCG@20 HR@10 NDCG@10
Baselines: GRU4Rec SASRec BERT4Rec S3-Rec CL4SRec CoSeRec ICLRec DuoRec MCLRec ICSRec LRD RLMRec
Builds on: ICLRec DuoRec CL4SRec SASRec LLM-ESR
Closest papers by meaning
- LLMEmb: Large Language Model Can Be a Good Embedding Generator for Sequential Recommendation (AAAI 2025 · similarity 0.89)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.88)
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.88)
- Review-Enhanced Hierarchical Contrastive Learning for Recommendation (AAAI 2024 · similarity 0.88)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.87)
- Augmenting Sequential Recommendation with Balanced Relevance and Diversity (AAAI 2025 · similarity 0.87)
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.