← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Meta data scientist interview with a meaty fraud classifier case study. The whole thing was basically one long quantitative problem with a few conceptual layers stacked on top. Felt more like applied ML product design than a typical DS screen.

Questions Asked (3)

Q1

You have a credit card fraud classifier deployed as a probability scorer. Given a base fraud rate of 0.20% across 1 million transactions, and asymmetric costs of $15 per false positive and $100 per false negative, compute the expected TP, FP, FN, TN counts and total cost for three candidate thresholds. Which threshold minimizes cost and what are the business trade-offs?

Product Analytics & MetricsTechnical Trade-offsA/B Testing & Experimentation
Author's notes

This is the kind of question where the math looks straightforward but you can fumble it fast if you mix up base rates.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify the threshold values and the classifier's score distribution (e.g., ROC or PR curve) to compute confusion matrices. Then, calculate expected TP, FP, FN, TN and total cost for each threshold using the given base rate and cost matrix. Finally, compare costs and discuss business trade-offs such as fraud prevention vs. customer friction and operational capacity.

Pro tip: Always state your assumptions about the classifier's performance (e.g., TPR and FPR at each threshold) and note that in practice you'd use validation data to estimate these; this shows you understand the gap between theoretical calculations and real-world deployment.

1. Clarify inputs and assumptions

Confirm the three threshold values and the classifier's true positive rate (TPR) and false positive rate (FPR) at each threshold. If not provided, assume plausible values or state that you would derive them from a validation set.

2. Compute confusion matrix counts

Calculate the number of actual fraud cases (0.20% of 1M = 2,000) and legitimate transactions (998,000). For each threshold, compute TP = TPR * 2,000, FN = 2,000 - TP, FP = FPR * 998,000, and TN = 998,000 - FP.

3. Calculate total cost per threshold

For each threshold, compute total cost = (FP * $15) + (FN * $100). Compare the costs to identify the threshold that minimizes total cost.

4. Analyze business trade-offs

Discuss the implications: a lower threshold catches more fraud (fewer FN) but increases false positives, causing customer friction and operational costs. A higher threshold reduces false positives but misses more fraud. Consider factors like customer lifetime value, manual review capacity, and regulatory requirements.

Key Points to Mention

  • Base rate and class imbalance: only 0.2% fraud, so even a low FPR can yield many false positives.
  • Cost asymmetry: false negatives are ~6.7x more costly than false positives, so the optimal threshold likely favors higher recall.
  • Threshold selection should be based on the cost curve, not just accuracy; the minimum cost threshold may not be the one with highest accuracy.
  • In practice, use a validation set to estimate TPR/FPR and consider precision-recall curves for imbalanced data.
  • Business trade-offs: customer experience (false positives block legitimate transactions), fraud losses, manual review workload, and potential regulatory fines.
  • Meta context: at scale (1M transactions), even small changes in threshold can have large absolute impacts on cost and user experience.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q2

How would you calibrate the model's probability scores, monitor for data or label drift over time, and periodically re-tune the operating threshold by segment such as country, merchant category, or transaction amount?

Product Analytics & MetricsTechnical Trade-offsRoot Cause Analysis
Author's notes

Blanked a little on the calibration part at first.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around a lifecycle: calibrate first, then monitor, then re-tune thresholds by segment. Emphasize that calibration and thresholding are separate concerns—calibration fixes probability reliability, while thresholding optimizes business metrics per segment. Use concrete methods and metrics, and tie everything back to production constraints and trade-offs.

Pro tip: Don't just list techniques—explain how you'd validate calibration (e.g., reliability diagrams, ECE) and set up automated drift alerts with clear ownership and retraining triggers. Mention that segment-level thresholds require sufficient sample sizes and guardrails against overfitting.

1. Calibrate probability scores

Apply post-hoc calibration methods like Platt scaling or isotonic regression on a held-out set, and evaluate with reliability diagrams, Brier score, and Expected Calibration Error (ECE). Choose the method based on data size and whether you need monotonicity.

2. Monitor for data and label drift

Track input feature distributions (PSI, KL divergence) and prediction distributions over time, and monitor label drift via delayed ground truth. Set up automated alerts with thresholds and dashboards, and define retraining triggers.

3. Re-tune operating thresholds by segment

For each segment (country, MCC, amount bucket), optimize the threshold to maximize a business metric (e.g., F1, profit) subject to constraints, using cross-validation and ensuring minimum sample size. Consider hierarchical or pooled thresholds for small segments.

4. Validate and deploy with guardrails

Run A/B tests or shadow deployments to measure impact, and include guardrail metrics (e.g., false positive rate, latency). Implement a rollback plan and monitor for segment-level performance regressions.

5. Automate periodic re-tuning

Schedule regular re-calibration and threshold re-tuning (e.g., monthly) or trigger by drift alerts. Use a pipeline that retrains, recalibrates, and re-evaluates thresholds, with human review for significant changes.

Key Points to Mention

  • Calibration methods: Platt scaling, isotonic regression, and evaluation metrics like ECE, Brier score, reliability diagrams.
  • Drift detection: PSI, KL divergence, KS test for features; prediction drift; label drift with delayed labels.
  • Segment-level thresholding: optimize for business metric per segment, handle small samples via pooling or hierarchical models.
  • Trade-offs: calibration vs. discrimination, threshold vs. ranking, segment granularity vs. sample size.
  • Productionization: automated monitoring, alerting, retraining triggers, and A/B testing with guardrails.
  • Business impact: tie thresholds to cost-sensitive metrics (e.g., fraud loss vs. customer friction) and communicate to stakeholders.

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.

Q3

What guardrail metrics would you set up to catch harmful side effects like a spike in declines for high-value customers, and how would you design an experiment to validate threshold changes without causing serious customer harm?

A/B Testing & ExperimentationProduct Analytics & MetricsStakeholder Management
Author's notes

Liked this part more.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining guardrail metrics that directly capture harmful side effects, such as decline rates for high-value customer segments, and set thresholds based on historical data and business impact. Then, design a phased experiment with a holdback group and sequential monitoring to validate threshold changes, ensuring early detection and mitigation of harm. Emphasize collaboration with stakeholders to align on risk tolerance and escalation procedures.

Pro tip: Frame guardrails as part of a broader 'trust and safety' framework for experiments, and propose a 'circuit breaker' mechanism that automatically halts the experiment if guardrails are breached, showing you prioritize customer protection over short-term gains.

1. Identify high-value segments and define guardrail metrics

Segment customers by value (e.g., lifetime value, engagement) and select metrics like decline rate, complaint rate, or churn risk that indicate harm. Set thresholds based on historical baselines and acceptable risk levels.

2. Design a controlled experiment with guardrail monitoring

Use a randomized controlled trial with a holdback group to measure the impact on guardrails. Implement sequential testing or group sequential design to allow early stopping if guardrails are breached.

3. Validate threshold changes safely via phased rollout

Test new thresholds in a small, low-risk population first (e.g., 1% of traffic) with enhanced monitoring. Gradually expand only if no harm is detected, using a pre-registered analysis plan.

4. Establish escalation and mitigation protocols

Define clear actions if guardrails are triggered, such as pausing the experiment, rolling back changes, or adjusting thresholds. Communicate these protocols to stakeholders in advance.

5. Monitor and iterate with stakeholder feedback

Continuously track guardrail metrics and gather input from product, engineering, and customer teams. Use insights to refine thresholds and experiment designs for future tests.

Key Points to Mention

  • Segmentation of high-value customers (e.g., by LTV, engagement) to tailor guardrails
  • Use of sequential testing or group sequential design for early harm detection
  • Setting thresholds based on historical data and business impact analysis
  • Phased rollout or holdback groups to minimize customer exposure
  • Automated circuit breakers or kill switches to halt experiments
  • Stakeholder alignment on risk tolerance and escalation procedures

AI-generated suggestions, not part of the candidate's original notes. May be inaccurate — verify before relying on them.