所属模块：
M1 · 数学与统计基础 (Mathematics & Statistics Fundamentals)| 专题分类：实验设计 (A/B) (实验设计 (A/B))| 难度等级：Medium
一、核心一句话结论 (One-Sentence Summary)
用实验前协变量（如历史同指标）做回归调整，扣掉可解释方差。
CUPED removes predictable pre-experiment user variance by subtracting linear projections: $Y^ = Y – theta (X – E[X])$, reducing metric variance by factor $(1 – rho^2)$ and slashing required sample sizes.*
二、核心考点要义 (Key Insights)
- 📌 方差降低比例 ≈ ρ²
- 📌 不引入偏差（协变量与处理独立）
- 📌 是工业界 A/B 的标准降方差手段
English Insights:
– Core Mechanism: Uses historical pre-experiment data $X$ as a control variate; since $E[X – E[X]] = 0$, the treatment effect estimate remains completely unbiased.
– Optimal coefficient: $theta^ = frac{text{Cov}(Y, X)}{text{Var}(X)}$; identical to the ordinary least squares slope of $Y$ on $X$.
– Variance reduction factor: $text{Var}(Y^) = text{Var}(Y)(1 – rho^2)$, where $rho = text{Corr}(Y, X)$.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$Y^{cv}=Y-theta(X-bar X),qquad theta=frac{mathrm{Cov}(Y,X)}{mathrm{Var}(X)}$$
CUPED（Controlled-experiment Using Pre-Experiment Data）的构造：取实验前的协变量 X（通常是同一用户的历史同指标，如实验前 7 天的点击率），构造调整后的指标 Y^cv=Y−θ(X−X̄)，其中 θ=Cov(Y,X)/Var(X) 是最优系数。为什么降方差：Var(Y^cv)=Var(Y)−Cov(Y,X)²/Var(X)=Var(Y)(1−ρ²)，即方差降低比例为相关系数的平方 ρ²。若历史指标与实验期指标相关系数为 0.7，则方差降低 49%——等效于样本量翻倍。为什么无偏：由于随机分流使 X 在两组间分布相同（E[X|T=1]=E[X|T=0]），减去 θ(X−X̄) 不改变两组期望之差，故 ATE 估计仍无偏。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Let $Y$ be the experiment metric and $X$ be a pre-experiment metric measured before user exposure (hence strictly independent of treatment assignment: $E[X_A] = E[X_B] = E[X]$). Define adjusted metric $Y^* = Y – theta(X – E[X])$. The treatment effect estimate is $hat{Delta}_{text{CUPED}} = bar{Y}_B^* – bar{Y}_A^* = (bar{Y}_B – bar{Y}_A) – theta(bar{X}_B – bar{X}_A)$. Taking expectations: $E[hat{Delta}_{text{CUPED}}] = E[bar{Y}_B – bar{Y}_A] – theta(E[X] – E[X]) = Delta$, preserving exact unbiasedness. The variance is $text{Var}(Y^*) = text{Var}(Y) + theta^2text{Var}(X) – 2thetatext{Cov}(Y, X)$. Differentiating with respect to $theta$ and setting to 0 gives optimal $theta^* = frac{text{Cov}(Y, X)}{text{Var}(X)}$. Substituting $theta^*$ back: $text{Var}(Y^*) = text{Var}(Y) – frac{text{Cov}(Y, X)^2}{text{Var}(X)} = text{Var}(Y)left(1 – frac{text{Cov}(Y, X)^2}{text{Var}(Y)text{Var}(X)}right) = text{Var}(Y)(1 – rho^2)$.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
实践要点：① 协变量的选择——首选实验前同指标（相关性最高）；也可用多个协变量的回归调整（CUPED 的推广，等价于用实验前数据预测实验期指标再取残差）；② 前提条件——协变量必须在随机化之前测量（否则可能被处理影响），且与处理独立；实践中还要注意协变量缺失（新用户无历史）时退化为原指标；③ 与其他降方差方法的对比——分层随机化/后分层（Post-stratification）在离散协变量上有效，CUPED 在连续协变量上更灵活且常更强；两者可叠加；④ 实现的坑——θ 必须用实验数据估计（可能引入微小偏差，可用样本分割消除）、协变量需做异常值处理（重尾协变量会削弱效果）、新用户群体需单独处理。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
Industrial impact: In e-commerce and streaming where user engagement is highly stable (e.g. Past 2-week spend predicts future 2-week spend with $rho approx 0.7$), CUPED variance reduction is $1 – 0.7^2 = 51%$. This cuts the required experiment duration and sample size in half! If a user is brand new (cold start with no pre-experiment history), set $X_i = bar{X}$ so $X_i – bar{X} = 0$, gracefully falling back to unadjusted $Y_i$.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 用实验期间的数据作为协变量（引入偏差）
- ⚠️ 忽略新用户无历史协变量的退化情形
English Pitfalls:
– Using covariates $X$ that are measured after the user received treatment (which can be influenced by treatment, causing post-treatment conditioning bias).
– Computing $theta$ separately for Treatment and Control groups (must compute a pooled $theta$ across both to preserve exact unbiasedness).
六、高频深度面试追问与预测 (Follow-Up Questions)
- CUPED 需要什么前提？
- Why is CUPED mathematically equivalent to ANCOVA (Analysis of Covariance) linear regression $Y = alpha + beta T + gamma X + epsilon$?
- 与分层抽样/后分层的区别？
- How does CUPAC (CUPED with Machine Learning Predictions) generalize pre-experiment feature sets using Gradient Boosted Trees?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
工业级 A/B 实验设计、分流正交、SRM 卡方排查与方差缩减(Industrial A/B Testing: Split, SRM & Variance Reduction) - 🗺️ 知识图谱模块：
数据科学与因果实验导图
🔬 算法科学家与机器学习深度考察全量题库 (Science Depth)
本题收录于 TalentMe 算法科学家深度考察真题库 (Science Depth)。全库共 856 道硬核考点，深度覆盖数学统计、经典ML、深度学习、Transformer、大语言模型、多模态、推荐系统与 MLOps。支持 Jev 面经智能匹配、一键离线单文件 HTML 手册导出并直连 Obsidian 本地记忆。