OneRec

传统级联推荐架构方案缺点

• [[Objective Collision]] 优化目标冲突
  • Conflicts from Diverse Objectives 多目标之间的冲突
  • Cross-Stage Modeling Conflicts 跨阶段建模冲突
• Lag Behind AI Evolution 落后于 AI 进化

[[Figure 2:Comparison between a cascaded recommender system and the OneRec.]]

模型架构

• [[OneRec Tokenizer]]
  • 直接使用 ID 存在的问题 #card
    • 太稀疏，每天新增，贡献关系少，很难学到上下文的语义
• [[Multi-Scale Feature Engineering]]
• [[OneRec Encoder]]
• [[OneRec Decoder]]
• [[OneRec Reward System]]
  • [[2.4.1 User Preference Alignment]]
  • [[2.4.2 Generation Format Regularization]]
  • [[2.4.3 Industrial Scenario Alignment]]
    • 通过将业务目标放到奖励系统中解决

Figure 8:The overall process of OneRec’s post-training, including continual pre-training and reinforcement learning.#card

网络回响