所属模块：
M7 · 检索、排序与推荐系统 (Retrieval, Ranking & RecSys)| 专题分类：离线评估与 OPE (Offline Evaluation & OPE)| 难度等级：Hard
一、核心一句话结论 (One-Sentence Summary)
OPE 需三大假设（无混淆/正性/SUTVA）；常见失败：正性违反（新策略选新物品）、混淆、干扰、非平稳。
Off-Policy Evaluation relies on three foundational causal assumptions—Unconfoundedness, Common Support (Positivity), and SUTVA; violations caused by unobserved features, zero-probability actions, interference, or non-stationarity trigger catastrophic evaluation failures.
二、核心考点要义 (Key Insights)
- 📌 无混淆：展示只依赖可观测特征
- 📌 正性/重叠：新策略的动作在历史中有非零展示概率
- 📌 SUTVA：无干扰；另有’非平稳’（分布随时间变化）
English Insights:
– Unconfoundedness (No Hidden Confounders): Action selection depends strictly on observed features logged in data: r indep a | x.
– Common Support / Positivity: Every action the candidate policy can propose must have had a strictly positive probability of being chosen by the logging policy: pi_0(a | x) > 0.
– SUTVA (No Interference): One user’s recommendation treatment does not alter or interfere with the reward observed by another user.
– Non-Stationarity: Temporal drift in user preferences or catalog inventory breaks the assumption that historical response distributions mirror future reality.
三、核心数学原理与机理推导 (Mathematical Principles & Derivation)
$$text{assumptions}: text{unconfoundedness}, text{positivity}, text{SUTVA};qquad text{violation}Rightarrowtext{fail}$$
数学机理：OPE 的三大假设——(1) 无混淆（unconfoundedness / ignorability）——’展示’（动作）只依赖可观测的特征（无隐藏的混淆变量）；违反的后果——估计有偏（因为’隐藏因素’同时影响展示与奖励）；例子——’高价值用户更可能被展示广告’（若’用户价值’未观测，则估计有偏）。(2) 正性/重叠（positivity / overlap）——新策略会选的每个动作在历史数据中都有非零的展示概率（π_old(a|x)>0）；违反的后果——(a) 无法估计（该动作无数据 → 外推）；(b) 权重爆炸（π_old→0）；这是最易违反的假设——新策略常’推荐历史上没出现过的物品’（如新物品、长尾）；检测——检查’新策略的动作在历史中的展示率’（若有动作的展示率为 0 → 违反）。(3) SUTVA（Stable Unit Treatment Value Assumption）——(a) 无干扰（no interference）——一个物品的反馈不受’其他被展示物品’影响；(b) 无隐藏变体——同一动作只有一个版本；违反的后果——估计有偏；例子——推荐列表的’位置’影响点击（同一物品在不同位置点击率不同 → ‘动作’不只是’展示哪个物品’，还包括’位置’）。(4) 非平稳（non-stationarity）——(a) 用户兴趣/物品质量/市场随时间变化；(b) 历史数据的分布与当前不同；(c) 违反的后果——估计不适用（历史不代表现在）；(d) 对策——(i) 用’近期数据’；(ii) 加’时间特征’；(iii) 监控漂移。其他失败模式——(a) 倾向模型错误（π_old 估计不准 → IPS 有偏）；(b) 模型错误（DR 的模型不准 → 修正项方差大）；(c) 奖励的定义问题（奖励延迟、稀疏）；(d) 特征缺失（历史日志未记录某些特征 → 无法建模）；(e) ‘动作空间大’（推荐的动作空间百万级 → 倾向极难估计）。检测与缓解——(a) 正性检查——统计’新策略动作的历史展示率’；(b) 重叠度——计算’有效样本量（ESS）’；(c) 倾向模型诊断（AUC/校准）；(d) 模拟实验（用合成数据验证 OPE 的准确性）；(e) 与 A/B 对比（最终验证）；(f) 限制动作空间（只在’重叠区’评估）；(g) 用’随机化数据’补充（打破混淆与正性问题）。实践建议——(a) 先检查假设（尤其正性）；(b) 保留随机化数据（小流量随机是’金标准’）；(c) 限制评估范围（只在重叠区）；(d) 倾向模型要诊断；(e) 与 A/B 对比（最终验证）；(f) 注意非平稳（用近期数据）；(g) 离线只做粗筛。度量——(a) 正性违反率；(b) ESS；(c) 倾向模型的 AUC/校准；(d) OPE vs A/B 的偏差。
📖 查看英文严格数学推导 (English Mathematical Derivation)
Mathematical & Causal Formulation: Foundational OPE Assumptions.
(1) Assumption 1: Unconfoundedness (Conditional Ignorability):
$${r(x, a)}_{a in mathcal{A}} perp A mid X$$
Failure Mode: An unobserved variable $U$ (e.g., user seasonal mood, external marketing campaign, sudden celebrity tweet) influenced which item the old policy recommended AND influenced user click probability. Because $U$ is missing from feature log $X$, the backdoor path $A leftarrow U to R$ remains open, inducing severe causal bias.
(2) Assumption 2: Common Support / Overlap (Positivity):
$$forall x in mathcal{X}, , a in mathcal{A}: quad pi_1(a mid x) > 0 implies pi_0(a mid x) ge epsilon > 0$$
Failure Mode: Candidate policy $pi_1$ proposes recommending newly ingested items or novel category crosses that the historical policy $pi_0$ strictly never explored ($pi_0(a mid x) = 0$). The importance ratio $frac{pi_1(a)}{pi_0(a)} = frac{>0}{0}$ is mathematically undefined. OPE cannot evaluate actions never taken in history.
(3) Assumption 3: SUTVA (Stable Unit Treatment Value Assumption):
$$R_i(A_1, A_2, dots, A_N) = R_i(A_i)$$
Failure Mode (Marketplace Interference): If policy $pi_1$ recommends a limited-inventory luxury hotel, User 1 booking it prevents User 2 from booking it. In two-sided ridesharing, dispatching driver to Rider 1 removes the driver from Rider 2. One user’s assignment directly corrupts another user’s potential outcome.
(4) Assumption 4: Stationarity (Distribution Stability):
$$P_{text{future}}(R mid X, A) = P_{text{past}}(R mid X, A)$$
Failure Mode: User preferences drift post-holiday or after competitor product launches.
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 ‘正性最易违反’是关键——新策略常推荐历史未展示的物品；面试中能指出这一点是深度理解的标志。② ‘位置是动作的一部分’——若’动作’只定义为’展示哪个物品’而忽略位置，则 SUTVA 违反（位置影响点击）。③ ‘随机化数据是金标准’——它能同时解决’混淆’与’正性’问题（因为随机策略对所有动作有正概率）。④ ‘非平稳’常被忽视——历史数据不代表现在；故需用近期数据。⑤ ‘倾向模型诊断’必需——π_old 估计不准则 IPS 有偏。⑥ 面试要点——被问’OPE 什么时候失效’，应给出’三大假设（无混淆/正性/SUTVA）+ 非平稳 + 其他失败（倾向错/模型错/奖励定义/特征缺失）+ 检测与缓解‘；能指出’正性最易违反’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 Deterministic logging policies break Positivity—if production runs a deterministic ranker ($a_t = argmax f(x)$), 99.9% of catalog items have $pi_0(a) = 0$ for a given context; Positivity is violated for virtually all candidate policies; reserving a continuous 2% traffic slice for randomized or epsilon-greedy logging is mandatory for OPE. ② Hidden confounders in multi-device sessions—a user browses shoes on desktop at work, then opens the mobile app at home and buys; logging mobile interactions without linking desktop session features introduces severe unobserved confounding; cross-device identity stitching closes the backdoor path. ③ Detecting Positivity violations—evaluating the fraction of candidate policy mass placed on zero-propensity actions: $alpha_{text{uncovered}} = mathbb{E}_{x}[sum_{a: pi_0(a mid x)=0} pi_1(a mid x)]$; if uncovered mass exceeds 5%, OPE must be rejected. ④ Synthetic actions and content generation—in generative recommendation or LLM prompt variations, the new policy outputs novel tokens never generated by the old policy; Positivity is 100% violated; OPE must rely on learned reward models (Direct Method) rather than importance sampling. ⑤ Marketplace cannibalization adjustments—when evaluating inventory-constrained systems, counterfactual simulators must model global inventory state transitions. ⑥ Interview takeaway—state the three core assumptions (Unconfoundedness, Positivity, SUTVA), detail how each assumption breaks in real-world recommender systems, explain why deterministic logging destroys Positivity, and outline diagnostic checks.
五、常见面试避坑陷阱 (Common Pitfalls & Traps)
- ⚠️ 不检查正性（新策略动作无历史数据）
- ⚠️ 忽略’位置是动作的一部分’（SUTVA 违反）
English Pitfalls:
– Attempting to perform importance sampling on candidate policies that propose new items, ignoring severe positivity violations (division by zero).
– Applying standard OPE to ride-sharing or marketplace inventory search without accounting for SUTVA violations caused by shared physical supply.
– Evaluating OPE across seasonal holiday boundaries (e.g., using Black Friday logs to evaluate a January shopping policy), violating stationarity.
六、高频深度面试追问与预测 (Follow-Up Questions)
- 哪个假设最易违反？
- How does randomized exploration traffic guarantee the Positivity (Common Support) condition for downstream offline policy evaluation?
- 违反假设时如何检测？
- What statistical tests detect the presence of unobserved confounders in historical recommendation click logs?
七、知识图谱对齐 (Knowledge Graph Anchor)
- 🔗 关联底层卡片：
推荐系统离线评估与离线策略评估 (OPE)：逆倾向得分 (IPS) 与重要性采样(Off-Policy Evaluation (OPE): IPS, Doubly Robust & Calibration) - 🗺️ 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。