Lecture 9 — Advanced Applications (高阶应用题)


Q1 — Advanced: Cross-Selling Uplift vs. Cannibalization(交叉销售提升 vs 自相残杀)

Question (EN): A retailer offers an add-on warranty at checkout. Historic data:

  • If a customer buys the premium laptop (event ), probability they also buy the warranty (event ) is .
  • If a customer buys the standard laptop (), .
  • Current mix: . A new policy nudges premium by showing a comparison card, raising to but reducing to because customers feel “already protected.” Should the retailer implement the policy if the expected warranty take-rate must not fall?

Q2 — Advanced: Supplier Quality Signal(供应商质量信号)

Question (EN): A buyer receives a “Pass/Fail” inspection signal () on batches from two suppliers (). Priors: . Signal characteristics: A batch fails. Should the buyer switch future orders to if “post-fail” defect risk from is proportional to ?


Q3 — Advanced: Fraud Queue Triage(欺诈分流优先级)

Question (EN): A bank’s fraud model flags 4% of transactions. True fraud rate . Model metrics: TPR , FPR . Investigators can process 2% of transactions daily. Should the top 2% scores form Tier-1 queue?


Q4 — Advanced: AB Test with Spillovers(A/B 溢出效应)

Question (EN): Open rate , . However, household sharing causes within-group correlation. Can we use a binomial test to compare rates?


Q5 — Advanced: Medical Screening Threshold(筛查阈值)

Question (EN): Disease prevalence . Test A: TPR , FPR (cost =0.85=0.00518). Goal: PPV ≥ 30% with lowest cost. Which test or sequence should be used?


Q6 — Advanced: Hiring Test Bias(招聘测试偏差)

Question (EN): Two positions use test to predict success . Overall . For candidates with internship (): ; for without internship (): . Should HR down-weight the test for group?


Q7 — Advanced: Marketing Lift Under Base-Rate Shift(基率漂移下的提升评估)

Question (EN): Last year: , . This year, natural base rate rises to , expected treated rate . Can we claim “same 3% lift”?


Q8 — Advanced: Inventory Recall Risk(召回风险)

Question (EN): Recall probability . If recall and undetected at factory, major claim occurs with . Detection TPR , FPR . Find monthly expected claim rate.


Q9 — Advanced: Bonus Plan Under No-Independence(非独立条件下的奖金计划)

Question (EN): Salespersons A and B have correlated success events: . If bonus pricing assumes independence, will budget be underestimated?


Q10 — Advanced: Vendor Blacklist Decision(供应商拉黑决策)

Question (EN): Report probability: , . Prior . Blacklist only if . Should we blacklist?


Q11 — Advanced: Early-Warning KPI(预警指标选择)

Question (EN): Project delay . Signals (missed meetings) and (risk items >5). Observed . If signals not conditionally independent, which threshold rule yields higher PPV: “any = 1” or “both = 1”?


Q12 — Advanced: Price-Match Policy Risk(价格保护政策)

Question (EN): Competitor discount event correlates with promotion week : . You refund price difference if competitor discounts. If you estimate cost using overall , is it biased?


Q13 — Advanced: Cold-Start Personalization(冷启动个性化)

Question (EN): New users () have base CTR . Model trained on old users (): . Can we directly transfer threshold to ?


Q14 — Advanced: Location Bias in NPS(NPS 的地点偏差)

Question (EN): High NPS stores are located in high-income areas (). If we ignore , results may be biased. How can probability framework avoid this?


Q15 — Advanced: Churn Save-Offer Targeting(流失挽留优惠)

Question (EN): Churn prior . Model’s top 10% have . Offer acceptance . If non-churners react negatively, how should we target?


Q16 — Advanced: Warranty Reserve Under Evidence Update(基于证据更新的保修准备)

Question (EN): Initial return rate . Alert : , . If occurs, compute posterior .


Q17 — Advanced: Safety Stock Under Joint Risk(联合风险下安全库存)

Question (EN): Two warehouses have correlated demand spikes . If safety stock computed as assuming independence, under positive covariance is it underestimated?


Q18 — Advanced: Campaign Attribution Paradox(归因悖论)

Question (EN): Clicks () and purchases () are correlated due to latent intent (). Thus overstates causal effect. How to obtain unbiased incremental lift?


Q19 — Advanced: Two-Stage Audit Policy(两阶段审计)

Question (EN): Fraud prior . Rule engine (R): TPR , FPR . ML model (M): on subset, TPR , FPR . Manual review only if both positive. What percent of all invoices are reviewed?

基率

③ 代入第二阶段 TPR/FPR 求总体阳性率。


Q20 — Advanced: Value of Negative Recommendation(负面建议的价值)

Question (EN): After a “negative” recommendation (), approval posterior . If preparing for relocation costs 500k if approved, what is optimal?