Maya (Customer Support Director) is upset.

Based on your metric choice in Q2, your model will flag approximately [Y] transactions per day for review. The fraud ops team can only handle 500 reviews/day.

Maya: "We're getting crushed with false positive complaints. Customers are furious when legitimate transactions get blocked."

What do you do?

Current tokens: [accumulated, keep track manually based on Q1, Q3, Q5, Q6]

The [Y] transactions depends on what you chose in Q2.

Precision: ~400 transactions/day
Recall: ~1,200 transactions/day
F1: ~700 transactions/day
Accuracy: ~65 transactions/day (dangerously low)

Question

Maya (Customer Support Director) is upset.

Based on your metric choice in Q2, your model will flag approximately [Y] transactions per day for review. The fraud ops team can only handle 500 reviews/day.

Maya: "We're getting crushed with false positive complaints. Customers are furious when legitimate transactions get blocked."

What do you do?

Current tokens: [accumulated, keep track manually based on Q1, Q3, Q5, Q6]

The [Y] transactions depends on what you chose in Q2.

Precision: ~400 transactions/day
Recall: ~1,200 transactions/day
F1: ~700 transactions/day
Accuracy: ~65 transactions/day (dangerously low)

BlackTom AI · Accepted Answer

Precision focuses on reducing false positives, while recall focuses on catching as many true frauds as possible. Key concepts: precision = true positives / (true positives + false positives); recall = true positives / (true positives + false negatives). Raising the model’s decision threshold increases precision (fewer flagged but more confident frauds) and reduces recall (more missed frauds).

Step-by-step reasoning:
1) Current operational constraint: fraud ops can review 500 transactions/day. Stakeholder Maya reports too many legitimate transactions being blocked (high false positives).
2) From the given metrics: a precision-optimized choice yields ~400 flagged/day (within capacity), recall-optimized yields ~1,200/day (overwhelms ops), F1 choice ~700/day (still >500), and accuracy choice ~65/day but with dangerously low model quality.
3) Increasing the precision threshold is the direct lever to reduce false positives immediately: it cuts the number of flagged transactions and raises the fraction that are true frauds. That directly addresses customer complaints and workload constraints.

Intermediate calculations/illustration:
- Recall option: ~1,200 flagged/day > 500 capacity → unsustainable.
- F1 option: ~700 flagged/day > 500 → still overloads team.
- Precision option: ~400 flagged/day ≤ 500 → workload manageable and fewer false positives.
- Accuracy option: ~65 flagged/day ≤ 500 but low accuracy implies many undetected frauds or unstable behavior.

Why the chosen action is correct:
- Raising the precision threshold costs zero operational resources and delivers immediate relief by reducing false positives and daily review volume to within 500/day.
- The trade-off is lower recall (some fraudulent transactions will be missed), which is why another stakeholder (Alex) is upset — they prioritize catching more fraud.

Why other options are less appropriate:
- Prioritizing recall or F1 floods the ops queue (>500/day) and worsens customer experience.
- Relying on accuracy masks class imbalance and doesn’t resolve false positives despite low flagged counts.

Conclusion: Increase precision threshold (Cost: 0) — Reduces recall, angers Alex. This matches the provided correct answer.

类似问题

Match each scenario (Column A) with the correct interpretation. 1: A model predicts all instances as positive. 2: A system flags very few positives, but most flagged cases are correct. 3: A model detects almost all actual positives but also produces many false alarms.

The confusion matrix is used to:

PayFlow has a 2.1% fraud rate (highly imbalanced). Customer Support is overwhelmed with false positive complaints, but Fraud Ops demands you catch every fraud case. Which metric will you optimize for? Current tokens: [based on Q1 choice, please track them yourself]

Which of the following are appropriate ways to evaluate the performance of a Support Vector Machine (SVM) used for classification? (Select all that apply)

Which indicators are used for model evaluation in classification? (Select all that apply)

更多留学生实用工具

智能学习助手

风格化写作助手

论文查重助手

文献引用助手

课堂转译助手

课堂笔记助手

Quiz搜索助手

学校历年真题

智能刷题助手

智能匹配练习

希望你的学习变得更简单