Research map / Large language models / Reasoning and chain of thought
CoT4Rec: Revealing User Preferences Through Chain of Thought for Recommender Systems
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 CoT4Rec, which uses GPT-3.5 to generate clustered chain-of-thought user preference descriptions, then trains a flan-T5-base preference analyzer and a flan-T5-base recommender to produce transparent two-stage sequential recommendations.
Problem
LLM-based recommenders suffer from uninterpretable discrete identifiers and implicit, hallucination-prone reasoning, leaving recommendation logic opaque.
Approach
A CoT reasoning strategy clusters sliding-window fragments of user histories (Sentence-BERT + k-means), uses representative sequences with zero-shot GPT-3.5 prompts as demonstrations to produce user preferences, then trains a small preference analyzer to generate these preferences without an external LLM, feeding them into a recommender that ranks candidates.
Key results
- CoT4Rec achieves N@10 0.4411 and HR@10 0.7295 on ML100K, outperforming the best baseline by 2.2%-12.2% across four datasets.
- On Electronics, CoT4Rec reaches N@5 0.3941 vs OpenP5 0.3746 (largest gains on Movies and Electronics exceed 5%).
- Ablation: removing the preference analyzer drops ML100K N@10 from 0.4411 to 0.4283.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoT4Rec instead of the strongest baseline, OpenP5 | Sequential recommendation · ML100K | N@10: 0.4157 -> 0.4411 | Helped |
| Use CoT4Rec instead of the strongest baseline, OpenP5 | Sequential recommendation · ML1M | N@10: 0.5479 -> 0.5676 | Helped |
| Use CoT4Rec instead of the strongest baseline, NCR | Sequential recommendation · Movies | N@10: 0.4670 -> 0.5158 | Helped |
| Use CoT4Rec instead of the strongest baseline, OpenP5 | Sequential recommendation · Electronics | N@10: 0.4054 -> 0.4329 | Helped |
| Add the user preference analyzer (PA) (ablation) | CoT4Rec · ML100K | N@10: 0.4283 -> 0.4411 | Helped |
| Add the user preference analyzer (PA) (ablation) | CoT4Rec · ML1M | N@10: 0.5244 -> 0.5676 | Helped |
| Add the user preference analyzer (PA) (ablation) | CoT4Rec · Movies | N@10: 0.4529 -> 0.5158 | Helped |
| Add the user preference analyzer (PA) (ablation) | CoT4Rec · Electronics | N@10: 0.3831 -> 0.4329 | Helped |
| Use the two-stage recommendation strategy (TS) (ablation) | CoT4Rec · ML100K | N@10: 0.4369 -> 0.4411 | Helped |
| Use the two-stage recommendation strategy (TS) (ablation) | CoT4Rec · ML1M | N@10: 0.5534 -> 0.5676 | Helped |
| Use the two-stage recommendation strategy (TS) (ablation) | CoT4Rec · Movies | N@10: 0.5122 -> 0.5158 | Helped |
| Use the two-stage recommendation strategy (TS) (ablation) | CoT4Rec · Electronics | N@10: 0.4211 -> 0.4329 | Helped |
Surprising findings
- Zero-shot GPT-3.5 CoT preferences performed worse than random manual demonstrations, showing the value of clustering-based demonstration selection.
Limitations
- Depends on external GPT-3.5 for generating training preference labels, with associated cost and hallucination risk.
- Evaluation restricted to sampled ranking (101 candidates) with leave-one-out on four datasets.
Experimental setup
Datasets: MovieLens ML100K MovieLens ML1M Amazon Movies Amazon Electronics
Benchmarks: MovieLens ML100K MovieLens ML1M Amazon Movies Amazon Electronics
Metrics: NDCG@5 NDCG@10 HR@5 HR@10
Baselines: BPR-MF SVD++ DMF NeuMF GRU4Rec STAMP DuoRec NCR OpenP5
Builds on: chain-of-thought prompting Auto-CoT OpenP5 NCR
Closest papers by meaning
- LLMRG: Improving Recommendations through Large Language Model Reasoning Graphs (AAAI 2024 · similarity 0.88)
- Bridging the User-side Knowledge Gap in Knowledge-aware Recommendations with Large Language Models (AAAI 2025 · similarity 0.88)
- CoRA: Collaborative Information Perception by Large Language Model’s Weights for Recommendation (AAAI 2025 · similarity 0.87)
- Reinforced Latent Reasoning for LLM-based Recommendation (ICLR 2026 · similarity 0.87)
- LLM-ESR: Large Language Models Enhancement for Long-tailed Sequential Recommendation (NeurIPS 2024 · similarity 0.86)
- Tree of Preferences for Diversified Recommendation (NeurIPS 2025 · 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.