M7-005Medium排序与推荐 · Ranking & recsys

冷启动问题怎么解决?新品/新用户/新场景分别怎么处理?

How do you solve cold start for new items, new users and new contexts?
💡 考场高分一句话答案:

新物品靠内容特征与探索,新用户靠人口统计与快速适应,新场景靠迁移学习与先验。

New items rely on content features and exploration, new users on demographics and fast adaptation, and new contexts on transfer from mature settings.

📌 核心考察要点 (Key Points)

  • 新物品:内容 embedding + 相似物品的协同信号;bandit 探索给初始曝光
  • 新用户:用注册信息/设备/地域作先验;前几次交互做快速个性化(contextual bandit)
  • 新场景:从成熟场景迁移(共享 embedding),或用规则兜底
  • 度量:新物品的“首次曝光→有效互动”转化率、探索预算占比、新用户 D1 留存
  • New items: content embeddings plus similarity to existing items; bandit exploration for initial exposure
  • New users: priors from signup, device and geography; rapid personalisation with a contextual bandit over the first few interactions
  • New contexts: transfer learning with shared embeddings, plus rule-based fallbacks
  • Metrics: first-exposure to first-engagement conversion, exploration budget share, D1 retention for new users

⚖️ 业务权衡与取舍 (Trade-offs & Calibration)

冷启动的本质是探索-利用权衡。纯利用会让新物品永远得不到曝光(马太效应),纯探索会伤害当期体验。实践做法是设“探索配额”(如 5% 曝光给新物品),并用 Thompson Sampling 或 UCB 分配,同时把探索成本计入护栏。

⚠️ 常见高频面试陷阱与避坑指南

  • 用全局平均作为新用户先验,导致所有人看到一样的内容
  • 探索预算无上限,伤害当期指标

🔍 大厂面试官连环追问 (Follow-ups)

  • Thompson Sampling 如何用于冷启动?
  • 如何衡量探索的长期价值?

嫌字多看太累?进入苏格拉底式对话推演实验室

零打字负担,以大厂考官真实困境切入,通过点选直觉选项暴露思维盲区,手把手推导 Why。

💬 开启对话式破冰 ➔
⚡
Test muscle memory on this topic via 3-Layer Flashcards
Active Recall: Mentally formulate Prompt ➔ Math Derivation ➔ Benchmark Implementation in 3 minutes.
🎯 Drill This Card Now➔