Research map / Label-efficient and robust learning / Recognition and long-tail
End-to-end Learnable Clustering for Intent Learning in 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 ELCRec, an end-to-end learnable clustering framework that jointly optimizes sequential recommendation and intent learning via learnable cluster centers and intent-assisted contrastive learning, with an industrial deployment on Alipay.
Problem
Existing intent learning methods alternate between clustering on full data (E-step) and representation learning (M-step), which limits scalability to large-scale data and yields sub-optimal performance.
Approach
Treat cluster centers (latent intents) as learnable network parameters trained with mini-batch pull-and-push clustering loss, and use these centers as self-supervision signals in intent-assisted contrastive learning, unifying behavior representation and clustering in one end-to-end objective.
Key results
- On Beauty, ELCRec improves NDCG@5 by 8.90% over the runner-up (0.0355 vs 0.0326) and reduces running time by 22.49% versus ICLRec.
- Averaged over four benchmarks, ELCRec reduces running time by 7.18% and GPU memory by 9.58% relative to ICLRec.
- In A/B testing on an Alipay live-streaming recommendation front page (130M PV), ELCRec improved PVCTR by 2.45% over baseline.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ICLRec with ELCRec | sequential recommendation · Sports, Beauty, Toys, Yelp | HR@5, HR@20, NDCG@5, NDCG@20: ELCRec exceeds ICLRec on all reported metrics except Toys HR@5 (0.0585 vs 0.0586); Toys HR@5 improvement: 0.17%↓ | Mixed |
| add ICL to the backbone (ablation) | backbone recommendation model · Sports, Beauty, Toys, Yelp | sum of HR@5, HR@20, NDCG@5, and NDCG@20: 0.1343 -> 0.1379; 0.239 -> 0.2398; 0.2664 -> 0.2675; 0.1258 -> 0.1262 | Helped |
| add ELCM to the backbone (ablation) | backbone recommendation model · Sports, Beauty, Toys, Yelp | sum of HR@5, HR@20, NDCG@5, and NDCG@20: 0.1343 -> 0.1396; 0.239 -> 0.2432; 0.2664 -> 0.2718; 0.1258 -> 0.1285 | Helped |
| add ICL to the backbone with ELCM (ablation) | B+ELCM · Toys | sum of HR@5, HR@20, NDCG@5, and NDCG@20: 0.2718 -> 0.2686 | Hurt |
| add ICL and ELCM to the backbone (ablation) | backbone recommendation model · Sports, Beauty, Toys, Yelp | sum of HR@5, HR@20, NDCG@5, and NDCG@20: 0.1343 -> 0.1405; 0.239 -> 0.2473; 0.2664 -> 0.2686; 0.1258 -> 0.1305 | Helped |
| replace ICLRec with ELCRec | industrial recommendation system · Alipay livestreaming and merchandise recommendation | PVCTR, VV, UVCTR: 2.45%↑; 2.28%↑; 2.41%↑; 1.62%↑ | Helped |
Surprising findings
- On the Toys dataset, adding intent-assisted contrastive learning to the ELCM variant lowered performance (ELCRec did not beat B+ELCM), which the authors attribute to ICL possibly degrading intent quality.
- On the Sports dataset, ELCRec's training time was slightly higher than ICLRec (5360 vs 5282 s), which the authors call a corner case.
Limitations
- Number of clusters (intents) k must be pre-defined and tuned per dataset.
- Hard to control the update rate of cluster centers, which the authors note as a drawback for rapidly shifting industrial data.
- Multiple loss weights (ICL weight fixed at 0.1, alpha tuned per dataset) require hyperparameter tuning; results not uniformly better across datasets (e.g., Toys).
Experimental setup
Datasets: Sports Beauty Toys Yelp ML-1M MIND-small
Benchmarks: Sports Beauty Toys Yelp ML-1M MIND-small
Metrics: HR@5 HR@20 NDCG@5 NDCG@20
Baselines: BPR-MF GRU4Rec Caser SASRec BERT4Rec DSSRec S3-Rec CL4SRec DCRec MAERec IOCRec ICLRec
Builds on: ICLRec S3-Rec SASRec CL4SRec
Closest papers by meaning
- De-collapsing User Intent: Adaptive Diffusion Augmentation with Mixture-of-Experts for Sequential Recommendation (AAAI 2026 · similarity 0.86)
- ATRank: An Attention-Based User Behavior Modeling Framework for Recommendation (AAAI 2018 · similarity 0.86)
- Intent Oriented Contrastive Learning for Sequential Recommendation (AAAI 2025 · similarity 0.86)
- Discovering New Intents with Deep Aligned Clustering (AAAI 2021 · similarity 0.85)
- Long-Sequence Recommendation Models Need Decoupled Embeddings (ICLR 2025 · similarity 0.84)
- Objective Shaping with Hard Negatives: Windowed Partial AUC Optimization for RL-based LLM Recommenders (NeurIPS 2026 · similarity 0.83)
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.