Files
myWiki/concepts/recommendation-reasoning.md

43 lines
1.7 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
title: "推荐推理 (Recommendation Reasoning)"
created: 2026-06-10
updated: 2026-06-10
type: concept
tags: [recommendation, reasoning, chain-of-thought]
sources: [raw/papers/onereason-team-onereason-2026.md]
---
# 推荐推理 (Recommendation Reasoning)
> 在推荐系统中引入「先思考再推荐」的推理范式,让模型在生成推荐结果前进行显式的逻辑推理。
## 核心挑战
与经典的数学/代码/符号逻辑推理不同,推荐推理具有三个独特性质:
1. **多解性**:推荐允许多个合理答案,非单一正确答案
2. **意图不可观测**:用户真实兴趣永远无法直接观测,只能从行为序列推断
3. **[[abductive-reasoning-recommendation|溯因推理]]**推荐推理是溯因Abduction而非演绎Deduction——从行为反推隐含兴趣
## 推理 vs 非推理的意外发现
初步研究OneRec-Think、OpenOneRec发现 **thinking mode 并不优于 non-thinking mode**。OneReason 论文将此归因于两个缺失要素:
- **[[perception-cognition-recommendation|Perception (感知)]]** 不足itemic token 缺乏语义 grounding
- **[[perception-cognition-recommendation|Cognition (认知)]]** 不足CoT 结构不适合推荐任务
## 推荐 CoT 的理想结构
一个好的推荐 CoT 应该:
1. 从行为序列中选择相关行为作为假设的兴趣点
2. 将兴趣压缩为可解释的偏好
3. 建模兴趣间的时间演化
4. 将推断的偏好状态与推荐 item 关联
## 参考
- [[onereason|OneReason]]
- [[recommendation-cot|推荐 CoT]]
- [[perception-cognition-recommendation|感知-认知推荐层次]]
- [[abductive-reasoning-recommendation|溯因推理 (推荐)]]