所属模块：
M2 · 经典机器学习 (Classical Machine Learning)| 专题分类：缺失值与数据泄漏 (Missing Values & Data Leakage)| 难度等级：Easy
一、核心一句话结论 (One-Sentence Summary)
MCAR（完全随机）、MAR（依赖可观测变量）、MNAR（依赖缺失值本身）。
Rubin’s taxonomy defines MCAR (completely random), MAR (conditional on observed data), and MNAR (dependent on unobserved missing values); treatments range from simple imputation to pattern modeling.
二、核心考点要义 (Key Insights)
- 📌 MCAR/MAR 可插补；MNAR 需建模缺失机制
- 📌 删除法在 MCAR 下无偏但损失信息
English Insights:
– MCAR: missingness independent of both observed and unobserved data; deletion is unbiased but loses power
– MAR: missingness depends on observed features; multiple imputation (MICE) and tree surrogate splits are valid
– MNAR: missingness depends on the missing value itself; requires explicit missing indicators or Heckman selection models
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{MCAR}: P(Rmid X_{obs},X_{mis})=P(R)$$
三种机制的严格定义（R 为缺失指示变量）：① MCAR（Missing Completely At Random）——缺失与任何变量（含缺失值本身）都无关，即 P(R|X_obs,X_mis)=P(R)；例如随机仪器故障导致的缺失。后果：完整案例分析（删缺失）无偏但损失信息；插补也无偏。② MAR（Missing At Random）——缺失依赖于可观测变量，即 P(R|X_obs,X_mis)=P(R|X_obs)；例如’收入’缺失与’年龄’相关（年轻人更可能不填），但给定年龄后缺失与收入无关。后果：删除会有偏，但基于可观测变量的插补（MICE、回归插补）可得到无偏估计。③ MNAR（Missing Not At Random）——缺失依赖于缺失值本身，即 P(R|X_obs,X_mis)≠P(R|X_obs)；例如’高收入者更不愿披露收入’。后果：最麻烦——标准插补都有偏，需显式建模缺失机制（选择模型、Heckman 校正、模式混合模型），且假设不可检验（因为缺失值不可观测）。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Let $Y = (Y_{text{obs}}, Y_{text{mis}})$ and $M$ be the binary missingness indicator ($M_i = 1$ if missing).
① MCAR (Missing Completely at Random): $P(M | Y_{text{obs}}, Y_{text{mis}}, X) = P(M)$. Deletion retains unbiased parameter estimates. ② MAR (Missing at Random): $P(M | Y_{text{obs}}, Y_{text{mis}}, X) = P(M | Y_{text{obs}}, X)$. Missingness is explained entirely by observable covariates (e.g., young people are less likely to report income, but conditional on age, missingness is random). Standard imputation methods (KNN, MICE, EM) are unbiased. ③ MNAR (Missing Not at Random): $P(M | Y_{text{obs}}, Y_{text{mis}}, X)$ depends on $Y_{text{mis}}$ directly (e.g., individuals with extremely high income systematically refuse to report income). Standard imputation is biased; requires joint modeling (Heckman selection model or pattern-mixture models).
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
处理策略与要点：① 机制判断——MCAR 可用 Little’s test 检验（检验缺失模式与可观测变量的独立性）；MAR 不可直接检验（只能基于领域知识）；MNAR 完全不可检验（需做敏感性分析：在不同 MNAR 假设下看结论是否稳健）。② 方法选择——MCAR：删除或简单插补均可；MAR：多重插补（MICE） 是标准方法（多次抽样反映不确定性，比单次插补更正确地反映方差）；MNAR：需专门建模或敏感性分析。③ 简单插补的问题——均值/中位数插补会压缩方差（所有缺失值取同一个数）并扭曲相关性（人为降低方差导致相关系数偏高），应避免用于统计分析；树模型可用代理分裂或默认方向原生处理缺失。④ 指示变量——当缺失本身有信息（MNAR 或缺失与目标相关）时，加入’是否缺失’的二值特征可显著提升模型（这是最实用的技巧之一）。⑤ 缺失比例——若某特征缺失 >50–70%，通常直接删除该特征；但需先确认缺失是否随机。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
Practical engineering strategies: In tabular ML, always append a Missing Indicator Column ($I_{text{missing}} in {0, 1}$) alongside continuous imputation. This preserves MNAR informative absence. Tree algorithms like LightGBM and XGBoost natively assign missing values to optimal split directions during training.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 对 MNAR 用均值插补（有偏且不可检验）
- ⚠️ 用均值插补后做统计分析（低估方差）
English Pitfalls:
– Using listwise deletion on MAR/MNAR data, introducing severe selection bias
– Imputing missing values with mean or median without adding a missing indicator, destroying MNAR signal
六、高频深度面试追问与预测 (Follow-Up Questions)
- MNAR 为什么最麻烦？
- How does XGBoost determine which branch a missing value should traverse during inference?
- 如何检验 MCAR？（Little’s test）
- What statistical test can be used to distinguish MCAR from MAR (e.g., Little’s MCAR test)?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
工业级缺失值填补与数据泄漏 (Data Leakage) 防范准则(Missing Value Imputation & Preventing Data Leakage) - 🗺️ 知识图谱模块：
机器学习工程师高频考点导图
🔬 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。