Research map / Label-efficient and robust learning / Recognition and long-tail
Preference Aware Dual Contrastive Learning for Item Cold-Start 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 PAD-CLRec, a cold-start recommender that adds a group-level preference-aware contrastive loss (aggregated warm-item collaborative features of a user vs. item content features) on top of item-level dual contrastive alignment.
Problem
Item cold-start recommendation: newly arriving items have no interactions, so their collaborative features are unavailable; existing item-level alignment methods ignore this condition during training, limiting performance especially for low-similarity cold items.
Approach
Treat a randomly selected warm item as cold during training and align its content feature with a user preference vector, computed by aggregating collaborative features of the user's other warm items, via a group-level preference-aware contrastive loss, combined with item-level dual contrastive losses in a joint objective.
Key results
- On Amazon Rec (All task), PAD-CLRec reaches Recall@10 2.20 vs 1.95 for the best baseline CLCRec (12.8% relative improvement).
- On Amazon Fashion (All task), PAD-CLRec reaches Recall@50 29.73 vs 28.75 for CLCRec (3.4% improvement).
- Ablation shows removing the preference-aware loss lowers Amazon Rec All Recall@20 from 3.22 to 3.04.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add the preference-aware contrastive loss to PAD-CLRec (ablation) | PAD-CLRec · Amazon Rec and Amazon Fashion, All/Warm/Cold | Recall@20: 3.04 -> 3.22; 27.41 -> 28.11; 3.91 -> 4.18; 25.18 -> 27.39; 4.51 -> 4.89; 30.11 -> 30.41 | Helped |
| Add the preference-aware contrastive loss to PAD-CLRec (ablation) | PAD-CLRec · Amazon Rec and Amazon Fashion, All/Warm/Cold | NDCG@20: 1.47 -> 1.51; 20.29 -> 21.01; 1.84 -> 1.97; 22.48 -> 23.95; 2.13 -> 2.24; 21.74 -> 27.12 | Helped |
| Add the dual terms in the recommendation and item-level alignment losses (ablation) | PAD-CLRec · Amazon Rec and Amazon Fashion, All/Warm/Cold | Recall@20: 3.15 -> 3.22; 26.72 -> 28.11; 4.00 -> 4.18; 27.27 -> 27.39; 4.49 -> 4.89; 28.10 -> 30.41 | Helped |
| Add the dual terms in the recommendation and item-level alignment losses (ablation) | PAD-CLRec · Amazon Rec and Amazon Fashion, All/Warm/Cold | NDCG@20: 1.40 -> 1.51; 17.67 -> 21.01; 1.73 -> 1.97; 19.37 -> 23.95; 2.07 -> 2.24; 24.63 -> 27.12 | Helped |
| Use Average rather than Weighted Sum for preference calculation (ablation) | PAD-CLRec · Amazon Rec and Amazon Fashion, All/Cold | Recall@20: 3.20 -> 3.22; 26.91 -> 28.11; 4.21 -> 4.18; 27.29 -> 27.39; 4.69 -> 4.89; 28.35 -> 30.41 | Mixed |
| Use Average rather than Weighted Sum for preference calculation (ablation) | PAD-CLRec · Amazon Rec and Amazon Fashion, All/Cold | NDCG@20: 1.47 -> 1.51; 19.92 -> 21.01; 2.22 -> 2.24; 21.61 -> 27.12 | Helped |
| Use Average rather than Sampling for preference calculation (ablation) | PAD-CLRec · Amazon Rec and Amazon Fashion, All/Cold | Recall@20: 3.08 -> 3.22; 27.05 -> 28.11; 3.91 -> 4.89; 28.54 -> 30.41 | Helped |
| Use Average rather than Sampling for preference calculation (ablation) | PAD-CLRec · Amazon Rec and Amazon Fashion, All/Cold | NDCG@20: 1.46 -> 1.51; 20.00 -> 21.01; 2.13 -> 2.24; 26.00 -> 27.12 | Helped |
Surprising findings
- The Average preference aggregation outperformed the Weighted Sum (softmax attention) variant on Cold and All tasks.
Limitations
- Large eta over-strengthens group-level alignment and harms cold and overall performance.
- Gains are modest on the denser Amazon Fashion dataset (around 1-3%).
Experimental setup
Datasets: Amazon Rec Amazon Fashion
Benchmarks: Amazon Rec Amazon Fashion
Metrics: Recall@K NDCG@K
Baselines: BPR DropoutNet MTPR GAR CFCCRec CLCRec
Builds on: CLCRec DropoutNet MTPR BPR
Closest papers by meaning
- Temporally and Distributionally Robust Optimization for Cold-Start Recommendation (AAAI 2024 · similarity 0.89)
- Counterfactual Task-augmented Meta-learning for Cold-start Sequential Recommendation (AAAI 2025 · similarity 0.85)
- Future Sight and Tough Fights: Revolutionizing Sequential Recommendation with FENRec (AAAI 2025 · similarity 0.84)
- SRA-CL: Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation (NeurIPS 2025 · similarity 0.84)
- Mult-DPO: Multinomial Direct Preference Optimization for Recommender Systems (NeurIPS 2026 · similarity 0.84)
- On Softmax Direct Preference Optimization for Recommendation (NeurIPS 2024 · similarity 0.84)
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.