M1-012Medium指标口径 · Metric definitions🔥 大厂高频出现 1 次

如何设计一个“用户满意度”指标?问卷和行为数据各有什么偏差?

How would you design a user satisfaction metric? What bias does survey data carry versus behavioural data?
💡 考场高分一句话答案:

行为代理做主指标(回访、举报、放弃),问卷做校准锚点;问卷有响应偏差与回忆偏差,行为有归因模糊的问题。

Use behavioural proxies as the primary metric and surveys as a calibration anchor; surveys carry response and recall bias, behaviour carries attribution ambiguity.

📌 核心考察要点 (Key Points)

  • 行为代理:D7 回访、举报率、卸载率、refinement rate、重复联系客服率
  • 问卷偏差:响应偏差(不满的人更可能答)、社会赞许偏差、回忆偏差、频率太低无法做实验
  • 设计要点:抽样而非全量弹窗(弹窗本身影响体验)、固定抽样比例、按分群加权还原总体
  • 组合方式:用问卷校准行为代理的方向,用行为代理做日常监控与实验
  • Behavioural proxies: D7 return, report rate, uninstall rate, refinement rate, repeat contact with support
  • Survey bias: non-response (the unhappy answer disproportionately), social desirability, recall, and too infrequent to power experiments
  • Design: sample rather than prompt everyone (prompting changes behaviour), keep the sampling rate fixed, weight by segment to recover the population
  • Combination: surveys calibrate the direction of the proxies; proxies carry daily monitoring and experiments

⚖️ 业务权衡与取舍 (Trade-offs & Calibration)

Google 的做法是把满意度调查当作“测量仪器”而非“产品功能”:固定 0.5–2% 的抽样率,长期不变,这样时间序列可比。一旦为了某次实验提高抽样率,历史对比就断了。

⚠️ 常见高频面试陷阱与避坑指南

  • 用弹窗问卷做主指标——弹窗会改变被测行为本身(观察者效应)
  • 忽略无响应者,把样本均值当总体满意度

🔍 大厂面试官连环追问 (Follow-ups)

  • 如何校正问卷的响应偏差?
  • 如果问卷满意度上升但留存下降,你信哪个?

🎯 关联大厂考题实战训练 (1 题)

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➔