Research map / Label-efficient and robust learning / Recognition and long-tail
Intent Oriented 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 IOCLRec, a sequential recommender that segments user histories into intent-level subsequences and trains three contrastive modules (fine-grained intent, single-intent, multi-intent) with intent-oriented augmentations.
Problem
Sequential recommendation models that encode the entire user history struggle to capture users' multiple, dynamically changing intents and their transitions.
Approach
Dynamically segment each user sequence into coarse-grained intent subsequences that serve as units for three contrastive objectives: K-means-based fine-grained intent contrast, single-intent augmentations (in-crop, in-mask, in-reorder), and multi-intent augmentations (in-insert, in-substitute) guided by item correlations and intent transitions.
Key results
- IOCLRec improves over the strongest baseline ICSRec by 10.28-18.68% on the three sparse Amazon datasets and 4.98-6.99% on ML-1M in HR and NDCG.
- Ablation on NDCG@10 shows removing the coarse-grained ICL module drops Beauty from 0.0657 to 0.0591 and ML-1M from 0.2048 to 0.1975.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IOCLRec instead of the strongest baseline, ICSRec | Sequential recommendation · Sports | HR@20: 0.0791 -> 0.0890 | Helped |
| Use IOCLRec instead of the strongest baseline, ICSRec | Sequential recommendation · Beauty | HR@20: 0.1295 -> 0.1504 | Helped |
| Use IOCLRec instead of the strongest baseline, ICSRec | Sequential recommendation · Toys | HR@20: 0.1365 -> 0.1620 | Helped |
| Use IOCLRec instead of the strongest baseline, ICSRec | Sequential recommendation · ML-1M | HR@20: 0.4345 -> 0.4541 | Helped |
| Add the coarse-grained intent contrastive learning loss (ablation) | IOCLRec · Sports, Beauty, Toys, ML-1M | NDCG@10: 0.0325 -> 0.0363; 0.0591 -> 0.0657; 0.0682 -> 0.0747; 0.1975 -> 0.2048 | Helped |
| Add the fine-grained intent contrastive learning loss (ablation) | IOCLRec · Sports, Beauty, Toys, ML-1M | NDCG@10: 0.0333 -> 0.0363; 0.0629 -> 0.0657; 0.0723 -> 0.0747; 0.2006 -> 0.2048 | Helped |
| Add the single-intent contrastive learning loss (ablation) | IOCLRec · Sports, Beauty, Toys, ML-1M | NDCG@10: 0.0342 -> 0.0363; 0.0611 -> 0.0657; 0.0696 -> 0.0747; 0.1993 -> 0.2048 | Helped |
| Add the multi-intent contrastive learning loss (ablation) | IOCLRec · Sports, Beauty, Toys, ML-1M | NDCG@10: 0.0347 -> 0.0363; 0.0628 -> 0.0657; 0.0704 -> 0.0747; 0.2012 -> 0.2048 | Helped |
| Replace intent-oriented augmentation operators with CoSeRec operators (ablation) | IOCLRec · Beauty and ML-1M | HR@20 | Hurt |
Surprising findings
- Gains are smaller on the dense ML-1M dataset, which the authors attribute to longer sequences with more diverse intents.
Limitations
- Gains are smaller on denser, longer-sequence ML-1M where intent diversity is higher.
- Many hyperparameters (sampling ratios, cluster count K, loss weights) must be tuned per dataset.
Experimental setup
Datasets: Amazon Sports Amazon Beauty Amazon Toys MovieLens-1M
Benchmarks: Amazon Sports Amazon Beauty Amazon Toys MovieLens-1M
Metrics: HR@10 HR@20 NDCG@10 NDCG@20
Baselines: BPR Caser SASRec BERT4Rec CoSeRec DuoRec ICLRec IOCRec ICSRec
Builds on: ICLRec ICSRec CoSeRec SASRec BERT
Closest papers by meaning
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.88)
- SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation (NeurIPS 2025 · similarity 0.86)
- End-to-end Learnable Clustering for Intent Learning in Recommendation (NeurIPS 2024 · similarity 0.86)
- Temporal Graph Contrastive Learning for Sequential Recommendation (AAAI 2024 · similarity 0.85)
- Augmenting Sequential Recommendation with Balanced Relevance and Diversity (AAAI 2025 · similarity 0.85)
- On the Unexpected Effectiveness of Reinforcement Learning for Sequential Recommendation (ICML 2024 · 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.