Research map / Large language models / Language, safety and interpretability
Tree of Preferences for Diversified 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 ToP-Rec, which uses an LLM to build a coarse-to-fine Tree of Preferences, infers users' underexplored preferences, and generates synthetic interactions to train a general recommender for diverse yet relevant recommendations.
Problem
Diversified recommendation suffers because observed user-item interactions are biased (exposure and selection bias), so underexplored user preferences are missed, which hurts diversity and relevance.
Approach
Use an LLM to construct a Tree of Preferences over items and reason top-down over it to uncover latent preferences; then match items to leaf preferences, synthesize interactions for underexplored preferences, and select influential users via a gradient-trajectory-based influence measure during training.
Key results
- ToP-Rec achieves the best Recall@50 (0.0586) and CE@50 (1.3275) on Twitter with LightGCN backbone, outperforming baselines in most cases.
- On Amazon, ToP-Rec attains R@50 0.1380 and CE@100 0.5902, the best in both metrics.
- Against DLCRec on Amazon at top-10, ToP-Rec improves R@10 from 0.0021 to 0.0119 and CE@10 from 0.4714 to 0.9639.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ToP-Rec instead of LightGCN | Recommendation on Twitter · Twitter | Recall@50: 0.0567 -> 0.0586 | Helped |
| Use ToP-Rec instead of LightGCN | Recommendation on Twitter · Twitter | Category-Entropy@50: 1.2841 -> 1.3275 | Helped |
| Use ToP-Rec instead of LightGCN | Recommendation on Weibo · Weibo | Recall@100: 0.1669 -> 0.1667 | Hurt |
| Use ToP-Rec instead of LightGCN | Recommendation on Weibo · Weibo | Category-Entropy@100: 1.0763 -> 1.1369 | Helped |
| Use ToP-Rec instead of LightGCN | Recommendation on Amazon · Amazon | Recall@50: 0.1362 -> 0.1380 | Helped |
| Use ToP-Rec instead of LightGCN | Recommendation on Amazon · Amazon | Category-Entropy@50: 0.5004 -> 0.5298 | Helped |
| Use ToP-Rec instead of DLCRec | Amazon recommendation · Amazon | Recall@10: 0.0021 -> 0.0119 | Helped |
| Use ToP-Rec instead of DLCRec | Amazon recommendation · Amazon | Category-Entropy@10: 0.4714 -> 0.9639 | Helped |
| Remove diversity contribution from item selection (ablation) | ToP-Rec · Twitter | Relevance | Mixed |
| Remove relevance contribution from item selection (ablation) | ToP-Rec · Twitter | Diversity | Mixed |
| Replace influential-user selection with random user selection (ablation) | ToP-Rec · Twitter | Relevance | Hurt |
| Replace influential-user selection with random user selection (ablation) | ToP-Rec · Twitter | Diversity | Hurt |
Surprising findings
- Heuristic baselines such as Random and MMR sometimes show diversity gains while losing recall, and LLM-based rerankers improve relevance but little diversity.
Limitations
- Effectiveness depends on quality of user and item textual attributes; weak for video or image platforms.
- Does not address novelty; only diversity is targeted.
Experimental setup
Datasets: Twitter Weibo Amazon
Benchmarks: Twitter Weibo Amazon
Metrics: Recall@50 Recall@100 Category-Entropy@50 Category-Entropy@100
Baselines: Random MMR DPP Box LCD-UC CDM LLM4Rerank-A LLM4Rerank-AD LLMRec-MMR DLCRec
Builds on: tree of thoughts BPR LightGCN MMR DPP
Closest papers by meaning
- Preference Diffusion for Recommendation (ICLR 2025 · similarity 0.88)
- Fading to Grow: Growing Preference Ratios via Preference Fading Discrete Diffusion for Recommendation (NeurIPS 2025 · similarity 0.87)
- DivGCL: A Graph Contrastive Learning Model for Diverse Recommendation (AAAI 2025 · similarity 0.87)
- On Softmax Direct Preference Optimization for Recommendation (NeurIPS 2024 · similarity 0.87)
- Robust Graph Based Social Recommendation Through Contrastive Multi-View Learning (AAAI 2025 · similarity 0.87)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
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.