← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Meta Data Scientist interview with a heavy stats and probability focus. The core problem was a multi-part fake account detection question that went pretty deep into Bayes, binomial distributions, and then got messier when they introduced signal correlation at the end.

Questions Asked (4)

Q1

You have a fake account classifier that uses five independent signals. Each signal fires with probability 0.8 for fake accounts and 0.05 for real ones. For a threshold of k=2 (flag if at least 2 signals fire), compute the probability of flagging a fake account and the probability of flagging a real account using the binomial distribution.

Product Analytics & MetricsTechnical Trade-offs
Author's notes

Binomial CDF complement, nothing exotic.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize that the number of signals firing follows a binomial distribution with n=5 and probability p (0.8 for fake, 0.05 for real). Compute the probability of at least 2 signals firing using the complement rule: 1 - P(X=0) - P(X=1). Then interpret these probabilities in the context of classifier performance, discussing trade-offs between false positives and false negatives.

Pro tip: After computing the probabilities, briefly mention how changing the threshold k would affect the trade-off between catching fakes and flagging real accounts, showing you understand the business implications beyond the math.

1. Identify the distribution

State that each signal is an independent Bernoulli trial, so the total number of signals firing follows a binomial distribution with n=5. Define p=0.8 for fake accounts and p=0.05 for real accounts.

2. Set up the probability calculation

We need P(X ≥ 2). Use the complement: P(X ≥ 2) = 1 - P(X=0) - P(X=1). Write the binomial formula for P(X=0) and P(X=1).

3. Compute for fake accounts

For p=0.8: P(X=0) = (0.2)^5 = 0.00032; P(X=1) = C(5,1)*(0.8)*(0.2)^4 = 5*0.8*0.0016 = 0.0064. So P(X ≥ 2) = 1 - 0.00032 - 0.0064 = 0.99328.

4. Compute for real accounts

For p=0.05: P(X=0) = (0.95)^5 ≈ 0.77378; P(X=1) = C(5,1)*(0.05)*(0.95)^4 = 5*0.05*0.814506 ≈ 0.20363. So P(X ≥ 2) = 1 - 0.77378 - 0.20363 = 0.02259.

5. Interpret the results

The classifier flags about 99.3% of fake accounts (true positive rate) and about 2.3% of real accounts (false positive rate). Discuss the implications for precision, recall, and potential threshold adjustments.

Key Points to Mention

  • Binomial distribution assumptions: independent signals, fixed number of trials (n=5), constant probability of firing for each class.
  • Complement rule for calculating P(X ≥ k) to simplify computation.
  • Exact calculations for both fake and real accounts, showing the steps.
  • Interpretation of results: high true positive rate (99.3%) and low false positive rate (2.3%).
  • Trade-offs: lowering k increases both true positives and false positives; raising k decreases both.
  • Business context: the cost of false positives vs. false negatives in fake account detection.

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

Q2

Given a base rate of 1.5% fake accounts and 5 million accounts scanned per day, use Bayes' theorem to find P(fake | flagged) and the expected number of flagged accounts. Is an 80,000-per-day manual review queue sufficient?

Product Analytics & MetricsA/B Testing & ExperimentationTechnical Trade-offs
Author's notes

This is where I slowed down.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify that the question requires additional information—specifically the false positive rate (FPR) and true positive rate (TPR) of the classifier—to compute P(fake | flagged). Then, set up Bayes' theorem with assumed or given values, calculate the expected number of flagged accounts, and compare it to the 80,000 review capacity. Finally, discuss the sufficiency of the review queue and potential trade-offs.

Pro tip: Always state your assumptions explicitly and note that in real-world scenarios, you would validate these assumptions with historical data or A/B tests. This shows rigor and awareness of practical constraints.

1. Identify missing parameters

Recognize that Bayes' theorem requires the true positive rate (sensitivity) and false positive rate (1 - specificity) of the flagging system. Without these, P(fake | flagged) cannot be computed.

2. Assume reasonable values

If not provided, assume plausible values for TPR and FPR based on industry standards or typical classifier performance (e.g., TPR = 0.9, FPR = 0.01). Clearly state these assumptions.

3. Apply Bayes' theorem

Compute P(fake | flagged) using the formula: P(fake | flagged) = (TPR * prevalence) / (TPR * prevalence + FPR * (1 - prevalence)). Use prevalence = 0.015.

4. Calculate expected flagged accounts

Multiply the total accounts scanned (5,000,000) by the probability of being flagged: P(flagged) = TPR * prevalence + FPR * (1 - prevalence). This gives the expected number of flagged accounts per day.

5. Assess review queue sufficiency

Compare the expected flagged accounts to the 80,000 daily review capacity. If expected flagged > 80,000, the queue is insufficient; otherwise, it may be sufficient. Discuss implications and possible adjustments.

Key Points to Mention

  • Bayes' theorem and its application to classification metrics
  • Importance of base rate (prevalence) in interpreting results
  • Trade-off between false positives and false negatives in content moderation
  • Operational impact of review queue capacity on user experience and safety
  • Need for validation of assumed TPR and FPR with real data
  • Potential strategies to improve precision or recall (e.g., threshold tuning, ensemble methods)

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

Q3

Find the smallest threshold k such that the expected daily flagged volume stays within 80,000 plus or minus 5%, while also maximizing P(fake | flagged). Walk through the trade-off between precision and recall.

Technical Trade-offsProduct Analytics & MetricsRoot Cause Analysis
Author's notes

This part took the most time.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the definitions and assumptions: what 'threshold k' means, how flagged volume and precision are computed, and the distribution of scores. Then, frame the problem as a constrained optimization: find the smallest k such that expected flagged volume is within [76k, 84k], and among those, maximize precision (P(fake|flagged)). Finally, discuss the precision-recall trade-off and how adjusting k affects both metrics.

Pro tip: Emphasize that in practice, you would validate the chosen threshold on a holdout set and monitor it over time, because the score distribution and base rates can drift. Also, mention that the 'smallest k' might not be unique if the volume constraint is an interval; you need to define whether you're minimizing k or maximizing precision within the interval.

1. Clarify definitions and assumptions

Define what 'threshold k' represents (e.g., a score cutoff), how flagged volume is computed (e.g., number of items with score >= k), and how precision is calculated. State assumptions about the score distribution and base rate of fakes.

2. Formulate the optimization problem

Express the goal as: minimize k subject to E[flagged volume] ∈ [0.95*80k, 1.05*80k], and then among feasible k, maximize P(fake|flagged). Note that if multiple k satisfy the volume constraint, choose the one with highest precision.

3. Analyze the precision-recall trade-off

Explain that increasing k reduces flagged volume (lower recall) but typically increases precision, because only higher-confidence items are flagged. Conversely, lowering k increases recall but decreases precision. The volume constraint forces a balance.

4. Solve and validate

Using historical data, compute the expected flagged volume and precision for candidate thresholds. Find the smallest k that meets the volume constraint, then check if a larger k within the constraint yields higher precision. Validate on a holdout set.

5. Discuss practical considerations

Mention that the optimal threshold may change over time due to distribution shifts, so monitoring and periodic re-evaluation are needed. Also, consider business impact: false positives vs. false negatives costs.

Key Points to Mention

  • Precision-recall trade-off: increasing threshold typically increases precision but decreases recall (and flagged volume).
  • Expected flagged volume constraint: need to estimate the distribution of scores to compute expected volume at different thresholds.
  • Base rate of fakes affects precision: even with high recall, precision can be low if fakes are rare.
  • The 'smallest k' may not be the one that maximizes precision; need to search within the feasible volume interval.
  • Validation on holdout data and monitoring for drift are essential for production.
  • Business context: the cost of false positives vs. false negatives should inform the choice of threshold.

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

Q4

If the five signals are not actually independent and have a pairwise correlation of 0.2 within each class, how does that change your analysis and what assumptions break down?

Technical Trade-offsData ModelingA/B Testing & Experimentation
Author's notes

Blanked for a moment.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify that correlation among signals reduces the effective information content, so the combined signal is less powerful than under independence. Then, quantify the impact on variance and power, and discuss how to adjust the analysis (e.g., using multivariate methods or correcting for correlation). Finally, address which assumptions break down and the practical implications for decision-making.

Pro tip: Mention that correlation can be addressed by decorrelating signals (e.g., PCA) or by explicitly modeling the covariance structure, but always validate with a simulation or empirical check. Also, note that even modest correlation can substantially inflate false positive rates if ignored in multiple testing.

1. Identify the broken assumptions

State that independence is violated, which affects variance calculations, combined signal distribution, and multiple testing corrections. Also, the effective number of independent tests is reduced.

2. Quantify the impact on variance and power

Show how correlation increases the variance of the average (or sum) of signals, reducing effective sample size and statistical power. For pairwise correlation ρ=0.2, the variance inflation factor is 1+(n-1)ρ for n signals.

3. Adjust the analysis method

Propose using multivariate techniques (e.g., MANOVA, mixed models) or decorrelation (PCA, whitening) to account for correlation. Alternatively, adjust p-values using methods that account for correlation (e.g., Benjamini-Yekutieli).

4. Re-evaluate decision thresholds

Recalculate critical values or thresholds for the combined signal to maintain the desired false positive rate. Consider simulation or permutation tests to empirically derive the null distribution.

5. Communicate practical implications

Discuss how the reduced power might affect experiment sensitivity and the risk of false negatives. Suggest ways to mitigate, such as increasing sample size or using more efficient estimators.

Key Points to Mention

  • Independence assumption is violated, affecting variance and distribution of combined signals.
  • Effective sample size is reduced; variance inflation factor for n signals with pairwise correlation ρ is 1+(n-1)ρ.
  • Multiple testing corrections (e.g., Bonferroni) become conservative; need methods that account for correlation.
  • Decorrelating signals via PCA or factor analysis can restore independence but may complicate interpretation.
  • Simulation or permutation tests can empirically estimate the null distribution under correlation.
  • Power loss may require larger sample sizes or more sensitive designs to detect true effects.

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