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：Hard
一、核心一句话结论 (One-Sentence Summary)
个体处理会影响他人（溢出），违反 SUTVA；用集群随机化、switchback 或专门的干扰感知设计。
Individual randomization fails under network interference because treating a user spills over to affect their connected social peers, violating SUTVA and contaminating control group baseline measurements.
二、核心考点要义 (Key Insights)
- 📌 干扰导致效应被稀释或放大（方向不定）
- 📌 干扰越强，个体随机化的偏差越大
English Insights:
– SUTVA Violation: In social networks (Meta, LinkedIn) or collaborative platforms (Slack, Figma), $Y_i$ depends not just on $T_i$, but on the treatment vector of friends $T_{mathcal{N}(i)}$.
– Direction of Bias: Direct network effects typically cause Treatment to elevate Control engagement (e.g. Chatting with untreated friends), severely diluting and underestimating the true treatment lift.
– Primary Mitigations: Graph Cluster Randomization (assigning tightly connected subgraphs to the same variant) and Ego-Network Randomization.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{SUTVA}: Y_i(T_i) text{独立于} T_j (jne i)$$
干扰的机制：标准 A/B 要求 SUTVA（个体 i 的结果只依赖自己的处理 Tᵢ），但在以下场景被违反——① 社交网络：处理组用户的行为会影响其好友（如分享、跟风），使对照组也被’间接处理’；② 双边市场：买家实验改变需求，影响卖家行为与另一侧用户体验；③ 共享资源：处理组占用更多库存/运力，挤压对照组（如打车平台的补贴实验）；④ 竞争/学习：同一关键词的广告预算实验会互相竞价。后果：个体随机化测到的是’直接效应 + 部分溢出效应’的混合——若溢出是正的（对照组也受益），则处理效应被低估；若是负的（对照组被挤压），则被高估。且干扰强度越大、偏差越大；严重时甚至符号翻转。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Network potential outcomes model (Manski, 2013): Let $Y_i(t, s)$ depend on individual treatment $t in {0, 1}$ and network exposure $s = frac{sum_{j in mathcal{N}(i)} T_j}{|mathcal{N}(i)|} in [0, 1]$. Under individual Bernoulli randomization, control users are exposed to $s > 0$. The observed difference in group means is: $hat{tau}_{text{naive}} = E[Y_i(1, s) mid T_i=1] – E[Y_i(0, s) mid T_i=0]$. However, the true total launch effect (Global Treatment Effect) is $tau_{text{global}} = E[Y_i(1, 1)] – E[Y_i(0, 0)]$. If positive spillovers exist ($E[Y(0, s)] > E[Y(0, 0)]$), then $hat{tau}_{text{naive}}$ systematically underestimates the true global product value.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
应对策略：① 集群随机化（Cluster Randomization）——以’无干扰的单元’为随机化单位（如地理区域、社交社区、城市）；优点是保留个体层面数据（功效高于 switchback）；代价是有效样本量降为集群数（集群内相关导致方差膨胀，需用 ICC（组内相关系数） 校正：设计效应 = 1+(m−1)·ICC），且集群数需足够（通常 ≥40）。② Switchback（时间轮换）——把干扰限制在时间窗内（见上题）；适合无法按空间划分的场景。③ 饱和设计（Saturation Design）——刻意让部分集群的处理比例不同（如 0%、50%、100%），通过对比不同饱和度的集群估计溢出效应；这是 Facebook 研究社交影响的标准方法。④ 双向随机化（Bipartite/Two-sided）——对市场两侧分别随机化并联合分析。⑤ 检测干扰——比较不同饱和度/集群大小的效应估计（若随饱和度变化则存在干扰）；或用’影子对照’（未暴露于任何处理的全隔离组）。⑥ 实践建议——先评估干扰强度（小规模集群实验），若干扰弱可用个体随机化 + 稳健标准误，若干扰强必须用集群/switchback；且应报告’干扰下的效应解读’（是直接效应还是总体效应）。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
Cluster Randomization protocol: (1) Run graph partitioning algorithms (Louvain or spectral clustering) to segment the social graph into isolated clusters minimizing cut edges between clusters. (2) Randomize entire clusters to Treatment or Control. This ensures friends share the same treatment status, containing spillovers. Trade-off: Reduces effective degrees of freedom from millions of individual users down to the number of clusters $K$, drastically inflating variance and requiring cluster-robust standard errors.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 在有网络效应的场景用个体随机化
- ⚠️ 集群随机化忽略 ICC 导致方差低估
English Pitfalls:
– Using individual-level standard errors on cluster-randomized data (drastically inflates false positive rate; must cluster standard errors by cluster ID).
– Allowing large inter-cluster boundary edge cuts, which lets spillovers bleed across cluster boundaries.
六、高频深度面试追问与预测 (Follow-Up Questions)
- 如何检测干扰的存在？
- How does the Exposure Model (Aronow & Samii, 2017) decompose network effects into direct and indirect spillover effects?
- 集群随机化的代价是什么？
- Why does graph modularity optimization (Louvain) maximize experimental power in network experiments?
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。