← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Meta DS interview with a heavy quant focus, basically one long fraud modeling case broken into three parts. The kind of question where you either live in this world daily or you're improvising the whole time.

Questions Asked (3)

Q1

You have a fraud scoring model outputting calibrated probabilities per account, with auto-block and manual review thresholds, a fixed daily review budget, and known costs for false positives, false negatives, and reviews. How do you write out the expected daily cost as a function of the two thresholds, and how would you estimate it from historical data using calibration techniques and empirical score distributions?

Product Analytics & MetricsTechnical Trade-offsData Modeling
Author's notes

This part took me longer than it should have to set up cleanly.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Define the expected daily cost as a sum over accounts of the expected cost given the two thresholds, using calibrated probabilities and the empirical score distribution. Then explain how to estimate this cost from historical data by binning scores, computing empirical rates, and applying calibration to ensure probabilities are reliable.

Pro tip: Emphasize that calibration is crucial because miscalibrated probabilities lead to suboptimal threshold choices; also mention that the review budget acts as a constraint, so the optimization is to minimize cost subject to the budget.

1. Define the cost function

Write the expected daily cost as a function of thresholds t1 (auto-block) and t2 (manual review), summing over all accounts the expected cost based on their calibrated fraud probability and the decision rule.

2. Incorporate the review budget constraint

Express the expected number of reviews as a function of t2 and the score distribution, and note that the budget imposes a constraint on t2 (or on the combination of t1 and t2).

3. Estimate from historical data

Use historical data to estimate the empirical score distribution (e.g., via histograms or kernel density estimation) and the calibration mapping (e.g., isotonic regression or Platt scaling) to obtain calibrated probabilities.

4. Compute expected cost empirically

For each score bin, compute the average calibrated probability and the empirical counts, then plug into the cost function to estimate the expected daily cost for any threshold pair.

5. Optimize and validate

Minimize the expected cost subject to the budget constraint, and validate the chosen thresholds on a holdout set or via cross-validation to avoid overfitting.

Key Points to Mention

  • Calibration techniques: isotonic regression, Platt scaling, reliability diagrams
  • Empirical score distribution: histogram, KDE, or quantiles
  • Cost matrix: false positive cost, false negative cost, review cost
  • Review budget as a constraint: expected number of reviews ≤ budget
  • Decision rule: auto-block if score > t1, manual review if t2 < score ≤ t1, allow if score ≤ t2
  • Optimization: minimize expected cost subject to budget, possibly using grid search or constrained optimization

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

Q2

Given the review budget constraint, how do you jointly optimize the two thresholds? Walk through how you'd pick an operating point on the precision-recall curve and then validate the choice with an online test.

A/B Testing & ExperimentationProduct Analytics & MetricsTechnical Trade-offs
Author's notes

The budget constraint is what makes this non-trivial.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a constrained optimization: maximize a business objective (e.g., expected value) subject to a review budget. Then explain how to select an operating point on the precision-recall curve using offline evaluation, and finally describe an online A/B test to validate the choice with guardrail metrics.

Pro tip: Emphasize that the optimal thresholds depend on the relative cost of false positives vs. false negatives and the review capacity; propose a sensitivity analysis to show robustness. Also, mention that you'd monitor for metric drift and re-optimize periodically.

1. Define the objective and constraints

Clarify the business goal (e.g., maximize precision at a fixed recall or maximize F-beta) and the review budget (e.g., number of items that can be manually reviewed per day). Translate the budget into a constraint on the number of positive predictions.

2. Model the precision-recall trade-off

Use offline validation data to plot the precision-recall curve for the model. Identify feasible operating points that satisfy the review budget constraint (i.e., predicted positive rate ≤ budget).

3. Select the operating point

Choose the point that maximizes the business objective (e.g., expected net benefit) while respecting the budget. Consider the cost matrix (cost of false positive vs. false negative) and perform sensitivity analysis.

4. Validate with an online experiment

Design an A/B test where the control uses current thresholds and the treatment uses the new thresholds. Define primary success metrics (e.g., precision, recall, revenue) and guardrail metrics (e.g., user experience, review workload).

5. Analyze and iterate

After sufficient sample size, analyze the results for statistical significance and practical significance. If successful, roll out; otherwise, re-evaluate and adjust thresholds.

Key Points to Mention

  • Precision-recall curve and how thresholds map to points on it
  • Review budget as a constraint on the number of positive predictions
  • Business objective function (e.g., expected value, F-beta score)
  • Cost matrix for false positives vs. false negatives
  • A/B testing methodology: randomization, sample size, guardrail metrics
  • Sensitivity analysis and monitoring for drift

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

Q3

Describe how you would monitor for model and score drift over time, and outline a weekly process for re-tuning the thresholds including backtesting and guardrails to prevent bad updates from going live.

Root Cause AnalysisSystem DesignTechnical Trade-offs
Author's notes

Felt most comfortable here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by defining the key metrics and data sources for monitoring model and score drift, then outline a weekly process that includes detection, root cause analysis, threshold re-tuning with backtesting, and guardrails for safe deployment. Emphasize automation, validation, and rollback mechanisms to ensure reliability.

Pro tip: Frame your answer around a real or hypothetical scenario where you caught drift early and prevented a bad update, showcasing your proactive monitoring and risk mitigation. Highlight the importance of aligning drift thresholds with business impact to avoid overreacting to noise.

1. Define Monitoring Metrics and Data Sources

Identify key performance indicators (e.g., AUC, calibration, score distribution) and data sources (e.g., prediction logs, ground truth labels) to track model and score drift. Establish baseline distributions and thresholds for alerting.

2. Implement Automated Drift Detection

Set up automated pipelines to compute drift metrics (e.g., PSI, KL divergence, KS test) on a daily/weekly basis. Use statistical process control or anomaly detection to flag significant deviations.

3. Conduct Root Cause Analysis

When drift is detected, investigate potential causes such as data quality issues, feature drift, concept drift, or upstream changes. Segment analysis by key dimensions to localize the problem.

4. Weekly Threshold Re-tuning with Backtesting

Re-tune decision thresholds using recent labeled data, and backtest on historical data to evaluate impact on key metrics. Simulate the new thresholds in a shadow mode before deployment.

5. Deploy with Guardrails and Rollback Plan

Use canary deployment or A/B testing to gradually roll out changes, with guardrail metrics (e.g., precision, recall, business KPIs) to automatically halt or rollback if degradation occurs. Document and review the process weekly.

Key Points to Mention

  • Types of drift: data drift, concept drift, and score drift, and how to differentiate them.
  • Statistical methods for drift detection: PSI, KL divergence, KS test, and their pros/cons.
  • Importance of ground truth labels and strategies for delayed labels (e.g., proxy metrics).
  • Backtesting methodology: time-based splits, walk-forward validation, and impact simulation.
  • Guardrails: canary releases, A/B testing, automated rollback triggers, and human-in-the-loop approval.
  • Business alignment: linking drift thresholds to business impact and cost of errors.

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