M1-017Hard指标树实例 · Metric trees

如何衡量推荐系统的“多样性”?多样性提升一定是好事吗?

How would you measure the diversity of a recommender? Is more diversity always better?
💡 考场高分一句话答案:

多样性要分“内容多样性 / 创作者多样性 / 消费多样性”三层度量,且必须与满意度联合看——多样性本身不是目标。

Measure at three levels — content, creator and consumption — and always read diversity jointly with satisfaction; diversity is not itself the objective.

📌 核心考察要点 (Key Points)

  • 度量:类目熵(entropy)、Gini 系数、distinct creator 数、intra-list similarity
  • 分层:曝光多样性(给了什么)、消费多样性(用户接受了什么)——前者容易刷,后者才反映真实偏好
  • 与满意度的关系:短期内多样性上升通常降低点击率,长期可能提升留存(避免审美疲劳)
  • 护栏:相关性不能崩、用户主动搜索/订阅行为不能被削弱
  • Metrics: category entropy, Gini, distinct creator count, intra-list similarity
  • Levels: exposure diversity (what was offered) versus consumption diversity (what was accepted) — the first is easy to inflate, only the second reflects real preference
  • Relationship to satisfaction: diversity usually lowers short-term CTR and may raise long-term retention by avoiding fatigue
  • Guardrails: relevance must not collapse, and active search or subscription behaviour must not weaken

📐 第一性原理与数理推导机理

类目熵 $H=-\sum_c p_c\log p_c$ 度量的是曝光分布的分散程度,但它对“把 1% 曝光给 100 个无关类目”也会给高分。因此需要与相关性联合:常用 $\text{Div}_{\text{adj}}=\frac{H}{\text{mean relevance}}$ 或在相关性约束下最大化熵。

⚖️ 业务权衡与取舍 (Trade-offs & Calibration)

多样性的真正价值在于探索(exploration):纯 exploit 会陷入局部最优并加剧马太效应。因此多样性指标应绑到“探索预算”上——如“每 20 条中有 2 条来自长尾”,这比抽象的熵更可执行、也更容易做实验。

⚠️ 常见高频面试陷阱与避坑指南

  • 只报熵上升就说“更健康”——熵高可能是推荐变差了
  • 忽略创作者侧的集中度(Gini),导致生态长期恶化

🔍 大厂面试官连环追问 (Follow-ups)

  • 如何设计一个“多样性 vs 相关性”的联合指标?
  • 多样性提升但 D7 留存下降,你怎么判断?

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➔