所属模块：
M7 · 检索、排序与推荐系统 (Retrieval, Ranking & RecSys)| 专题分类：在线指标与实验 (Online Metrics & Guardrails)| 难度等级：Medium
一、核心一句话结论 (One-Sentence Summary)
用户间/物品间相互影响（社交、供需、竞争）破坏 SUTVA；需集群随机化、switchback、或预算分割。
Network effects and shared marketplace resources violate the Stable Unit Treatment Value Assumption (SUTVA); treatment users interfere with control users through social contagion or inventory cannibalization, requiring cluster randomization, switchback tests, or synthetic controls.
二、核心考点要义 (Key Insights)
- 📌 干扰：一个用户的处理影响另一个用户（社交/竞争/供需）
- 📌 后果：A/B 的’处理效应’估计有偏（违反 SUTVA）
- 📌 对策：集群随机化、switchback、预算分割、地理分割
English Insights:
– SUTVA definition: Assumes the potential outcome of any experimental unit is independent of the treatment assignments of all other units.
– Two-sided marketplace interference: In Uber, Airbnb, or Amazon, treatment users booking rides or buying limited stock cannibalizes availability for control users.
– Social network spillover: In social feeds, treatment users sharing more content directly alters the feed experience of their control group friends.
– Mitigation architectures: Cluster-based randomization (graph community partitioning), Switchback testing (time-slice randomization), and Synthetic Controls.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{SUTVA violated}: text{units interact};qquad text{fix}: text{cluster randomization}, text{switchback}$$
数学机理：网络效应与干扰（interference）——(1) 定义——A/B 实验假设’一个单元的处理不影响其他单元’（SUTVA）；但很多场景违反：(a) 社交网络——’给 A 组推荐某内容 → A 组用户分享给 B 组用户 → B 组也受影响’；(b) 供需/市场——’给 A 组展示某商品 → A 组买走库存 → B 组买不到’（市场均衡效应）；(c) 竞争/拍卖——’A 组提高出价 → 广告价格上升 → B 组成本上升’；(d) 内容生态——’A 组看某创作者 → 该创作者获得更多曝光 → 影响 B 组看到的内容’；(e) 共享资源——’A 组消耗算力 → B 组延迟上升’。(2) 后果——(a) 处理效应估计有偏——’A 组的表现’包含了’对 B 组的溢出’（spillover）；(b) 方向不确定——溢出可能是正（社交传播）或负（竞争）；(c) ‘稀释’效应——若 A 组占比小，则’对整体的影响’被稀释（但 A/B 只测 A 组 vs B 组）；(d) ‘均衡效应’——大规模上线后的效果 ≠ 小流量实验的效果（因为’市场会调整’）。(3) 对策——(a) 集群随机化（cluster randomization）——把’相互影响的单元’分到同一组（如’同一社交圈的用户分到同一组’）；优点——减少组间干扰；缺点——(i) 需知道’谁与谁相互影响’；(ii) 有效样本量下降（因为集群内的单元’不独立’）。(b) Switchback 实验——时间上交替（如’第 1 小时用 A、第 2 小时用 B’）；适用——(i) 市场均衡场景（供需）；(ii) 无法按用户分割的场景（如’整个城市的定价’）；优点——避免’组间干扰’（因为同一时间只有一种处理）；缺点——(i) 需处理’时间趋势’（如’早晚高峰’）；(ii) 样本量受限（时间点数）。(c) 预算/流量分割——(i) 地理分割（不同城市不同策略）；(ii) 时间分割；(iii) 随机化单元上移（把’市场’作为单元）。(d) 建模干扰——(a) 显式建模’溢出效应’（如用’邻居的处理状态’作为特征）；(b) 用’因果推断的干扰方法’（如’暴露映射’）。(e) ‘小流量 + 大规模’的差异——(i) 小流量实验测’局部效应’；(ii) 大规模上线有’均衡效应’；(iii) 对策——分阶段放量，观察’效应是否随流量变化’。与其他问题的关系——(a) 与’SUTVA’（OPE 的假设）；(b) 与’长期效应’（均衡效应需长期观察）；(c) 与’实验平台’（需支持集群/switchback）。实践建议——(a) 先判断是否有干扰（社交/供需/竞争）；(b) 有干扰 → 集群随机化或 switchback；(c) 地理/时间分割（市场均衡）；(d) 建模溢出（显式）；(e) 分阶段放量（观察效应随流量的变化）；(f) 报告’有效样本量’（集群随机化会降低）。度量——(a) 溢出效应的估计；(b) 分阶段放量的效应变化；(c) 集群内的相关性（ICC）；(d) 有效样本量。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Causal & Experimental Methodology: Interference & SUTVA Violations.
(1) The Stable Unit Treatment Value Assumption (SUTVA):
Let $mathbf{W} = (W_1, W_2, dots, W_N) in {0, 1}^N$ be the global vector of treatment assignments. SUTVA requires:
$$Y_i(mathbf{W}) = Y_i(W_i) quad forall i in [1, N]$$
The potential outcome $Y_i$ depends only on user $i$’s own treatment $W_i$, completely invariant to how any other user $j$ is treated.
(2) Mechanisms of SUTVA Violations:
– Marketplace Supply Cannibalization (Indirect Interference):
Suppose Treatment algorithm recommends discounted hotel rooms. Treatment users book 80% of rooms. Control users search for rooms, find them sold out, and bounce ($Y_{text{control}} downarrow$). The estimated treatment effect $hat{tau} = bar{Y}_{text{treatment}} – bar{Y}_{text{control}}$ artificially inflates, falsely attributing control group starvation to treatment brilliance.
– Social Contagion (Direct Graph Spillover):
Treatment user $i$ is given a viral sharing feature; they share 10 posts. Friend $j$ (assigned to Control) sees these posts in their feed, increasing their engagement ($Y_{text{control}} uparrow$). The estimated lift $bar{Y}_T – bar{Y}_C$ collapses toward zero.
(3) Methodological Solutions:
– Cluster-Based Graph Randomization:
Partition the user social graph into disjoint communities $mathcal{C}_1, dots, mathcal{C}_K$ using graph clustering algorithms (e.g., Infomap / Louvain). Randomize entire clusters as atomic units ($W_c in {0, 1}$), bounding 95%+ of interference within cluster borders.
– Switchback Testing (Time-Slice Randomization):
For physical supply-demand platforms (Uber, DoorDash). Partition time into discrete windows (e.g., 30-minute intervals). The entire city alternates between Treatment and Control: $[T_1 = text{Control}, T_2 = text{Treatment}, T_3 = text{Control}, dots]$. Completely eliminates user-to-user supply cannibalization.
四、工业级落地权衡与工程考量 (Industrial Trade-offs)
深度剖析与工程权衡：① ‘干扰违反 SUTVA’是核心——面试中能指出’社交/供需/竞争’三类场景是深度理解的标志。② ‘Switchback 适合市场均衡’——时间上交替（同一时间只有一种处理）；这是’无法按用户分割’时的解法。③ ‘集群随机化降低有效样本量’——因为集群内单元不独立（ICC）；故需更大样本。④ ‘均衡效应使小流量与大规模不一致’——需分阶段放量观察。⑤ ‘建模溢出’是高级手段——显式建模’邻居的处理状态’。⑥ 面试要点——被问’什么场景不能直接 A/B’，应给出’社交/供需/竞争三类干扰 + 对策（集群随机化/switchback/地理分割/建模溢出）+ 分阶段放量‘；能指出’Switchback 适合市场均衡’是深度理解的标志。
⚙️ 查看英文落地权衡分析 (English Systems & Trade-offs)
In-Depth Analysis & Engineering Trade-offs: ① Switchback testing trade-offs—switchbacks eliminate marketplace cannibalization, but suffer from carryover effects: a surge in ride requests generated during Treatment period $T_2$ depletes driver supply at the start of Control period $T_3$; introducing 5-minute washout / buffer windows between switchback intervals mitigates carryover. ② Cluster randomization variance penalty—randomizing by clusters reduces sample size from $N$ users to $K$ clusters (e.g., 50 metropolitan regions); statistical variance increases dramatically, demanding larger effect sizes or multi-month experiment durations. ③ Synthetic Control Methods (SCM)—when testing platform-wide algorithmic changes across entire cities or countries (where only 1 or 2 markets receive treatment), synthetic controls construct a weighted combination of un-treated cities to form an artificial counterfactual baseline. ④ Detecting interference in standard user-split tests—tracking whether control group behavior changes as the treatment traffic allocation ramps from 5% to 50%; if control metrics drift as treatment volume expands, interference is actively contaminating the experiment. ⑤ Budget-split experiments in ad auctions—in ad bidding, treatment and control advertisers compete for the same query auctions; partitioning ad campaign budgets rather than user impressions isolates auction dynamics. ⑥ Interview takeaway—define SUTVA formally, explain supply cannibalization and social spillover, detail Cluster Randomization (for social graphs) and Switchback Testing (for two-sided physical marketplaces), and describe washout windows.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 在强干扰场景直接按用户随机（估计有偏）
- ⚠️ 小流量实验的效应直接外推到全量（均衡效应）
English Pitfalls:
– Running standard user-split A/B tests in two-sided marketplaces (ride-sharing, food delivery, hotel booking), severely inflating treatment effects via supply cannibalization.
– Deploying switchback testing without washout buffer intervals, allowing carryover supply depletion to poison adjacent control periods.
– Ignoring social graph network spillover, concluding a viral feature failed because control group friends also increased their activity.
六、高频深度面试追问与预测 (Follow-Up Questions)
- 哪些场景有强干扰？
- How do washout periods (buffer intervals) in switchback experiments prevent temporal carryover contamination between treatments?
- 什么是 switchback 实验？
- What graph partitioning objectives (such as minimizing normalized edge cuts) optimize cluster randomization in social networks?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
在线推荐实验与业务指标：CTR、CVR、留存时长、网络溢出效应与 CUPED(Online Metrics & A/B Testing: CTR, CVR, CUPED & Spillover) - 🗺️ 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。