← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
Jul 2026

Summary

Meta data scientist interview with a probability/Bayes question that spiraled into five parts. The core setup was straightforward but the later parts tripped me up more than I expected.

Questions Asked (5)

Q1

Given a pool of reviewers where 20% are lazy (always give good reviews) and 80% are careful (give good reviews 60% of the time), what is the probability that a randomly selected review is good?

Product Analytics & Metrics
Author's notes

This part was fine.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize this as a problem of total probability and use Bayes' theorem or a probability tree. Calculate the weighted average of the probabilities of a good review from each reviewer type, using the given proportions as weights. Clearly state the assumptions and walk through the calculation step by step.

Pro tip: After computing the probability, mention how this simple model could be extended to incorporate more reviewer types or varying probabilities, showing you think beyond the immediate question. Also, relate it to practical implications like estimating true quality from reviews.

1. Define the events and given probabilities

Let L be the event that a reviewer is lazy, C for careful, and G for a good review. Write down P(L)=0.2, P(C)=0.8, P(G|L)=1, P(G|C)=0.6.

2. Apply the law of total probability

Use the formula P(G) = P(G|L)P(L) + P(G|C)P(C) to combine the conditional probabilities with the prior probabilities of each reviewer type.

3. Compute the result

Plug in the numbers: P(G) = (1)(0.2) + (0.6)(0.8) = 0.2 + 0.48 = 0.68. So the probability is 68%.

4. Interpret and discuss implications

Explain that 68% of reviews are good on average. Optionally, discuss how this could be used to infer the proportion of lazy reviewers if only review outcomes are observed, or how the model might change with different assumptions.

Key Points to Mention

  • Law of total probability
  • Conditional probability
  • Weighted average
  • Assumption of independence between reviewer type and review outcome given type
  • Potential application to Bayesian inference or A/B testing
  • Clear communication of assumptions and limitations

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

Q2

If a review turns out to be negative, what is the probability it came from a lazy reviewer?

Product Analytics & Metrics
Author's notes

Bayes.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Clarify that this is a conditional probability problem requiring Bayes' theorem, and that we need to define 'lazy reviewer' and estimate base rates. Walk through the calculation using hypothetical but reasonable numbers, and discuss how the result depends on assumptions. Emphasize that in a real Meta context, you would validate these assumptions with data and consider business implications.

Pro tip: Don't just compute a number—frame the answer around how you would estimate the necessary probabilities from data (e.g., reviewer behavior logs, review sentiment) and how the insight would drive product decisions like reviewer incentives or quality checks.

1. Define terms and set up the problem

Define what 'lazy reviewer' means (e.g., a reviewer who doesn't put effort into writing detailed or fair reviews) and what 'negative review' means (e.g., low star rating or negative sentiment). State that we need P(Lazy | Negative).

2. Identify required probabilities

List the components needed for Bayes' theorem: P(Lazy), P(Negative | Lazy), P(Negative | Not Lazy), and P(Not Lazy). Explain how you might estimate these from historical data or experiments.

3. Apply Bayes' theorem

Use the formula P(Lazy | Negative) = P(Negative | Lazy) * P(Lazy) / P(Negative), where P(Negative) = P(Negative | Lazy)*P(Lazy) + P(Negative | Not Lazy)*P(Not Lazy). Plug in hypothetical values to illustrate.

4. Interpret and discuss sensitivity

Discuss how the result changes with different assumptions (e.g., if lazy reviewers are more likely to leave negative reviews). Highlight that the probability is highly dependent on the base rate of lazy reviewers and the likelihood of negative reviews from each group.

5. Connect to business context

Explain how this analysis could inform Meta's review moderation or reviewer quality initiatives, such as identifying lazy reviewers or improving review quality.

Key Points to Mention

  • Bayes' theorem and conditional probability
  • Base rate fallacy and importance of prior probabilities
  • Estimating probabilities from data (e.g., reviewer behavior, review sentiment)
  • Sensitivity analysis and assumption checking
  • Business implications for product analytics (e.g., reviewer incentives, quality control)
  • Clear communication of assumptions and limitations

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

Q3

What is the expected number of good reviews out of 100 total reviews?

Product Analytics & Metrics
Author's notes

Straightforward once you have part (a).

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Clarify the assumptions: define what constitutes a 'good review' (e.g., rating ≥4 stars) and the probability of a good review based on historical data or product context. Then model the number of good reviews as a binomial random variable with n=100 and probability p, and compute the expected value as 100p. If p is unknown, discuss how to estimate it from data or state that the expected number depends on p.

Pro tip: Always state your assumptions explicitly and connect the metric to business impact—e.g., 'If the current good review rate is 80%, we expect 80 good reviews, but improving it by 5% could yield 5 more good reviews, which might increase conversion by X%.' This shows you think beyond the math.

1. Clarify the definition

Ask or state what 'good review' means (e.g., 4+ stars, positive sentiment) and confirm that reviews are independent and identically distributed.

2. Identify the probability

Determine the probability p of a good review from historical data, A/B tests, or product benchmarks. If unknown, discuss how to estimate it.

3. Apply the binomial model

Model the number of good reviews as Binomial(n=100, p). The expected value is simply 100p.

4. Compute and interpret

Calculate the expected number (e.g., if p=0.8, expected good reviews = 80) and discuss variability (standard deviation = sqrt(100p(1-p))).

5. Connect to business context

Explain how this metric impacts business decisions, such as product improvements or marketing strategies, and suggest ways to increase p.

Key Points to Mention

  • Binomial distribution assumptions: fixed number of trials (100), independent reviews, constant probability p.
  • Expected value formula: E[X] = n * p = 100p.
  • Estimation of p from historical data or industry benchmarks.
  • Variance and standard deviation to understand uncertainty: Var(X) = 100p(1-p).
  • Business implications: how improving p affects the number of good reviews and downstream metrics.
  • Potential pitfalls: non-independence (e.g., viral effects), changing p over time, or selection bias in reviews.

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

Q4

A reviewer gives three consecutive good reviews. What is the probability that reviewer is lazy?

Product Analytics & MetricsAlgorithms & Data Structures
Author's notes

This is where it got interesting.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Clarify that this is a Bayesian inference problem: we need to estimate the probability that a reviewer is lazy given three consecutive good reviews. Define 'lazy' as a reviewer who tends to give positive reviews without much scrutiny, and set up prior probabilities and likelihoods for lazy vs. diligent reviewers. Then apply Bayes' theorem to compute the posterior probability, discussing assumptions and potential data needed.

Pro tip: Acknowledge that the answer depends heavily on the base rate of lazy reviewers and the probability of good reviews from each type; without these, the question is underspecified. Show how you would estimate these from historical data or make reasonable assumptions.

1. Define terms and assumptions

Define what 'lazy' means in this context (e.g., a reviewer who gives positive reviews with high probability regardless of quality) and what constitutes a 'good review'. State assumptions about independence of reviews and the prior probability of a reviewer being lazy.

2. Set up Bayesian model

Let L be the event that the reviewer is lazy, and G be the event of three consecutive good reviews. Specify prior P(L), and likelihoods P(G|L) and P(G|not L). Use Bayes' theorem: P(L|G) = P(G|L)P(L) / [P(G|L)P(L) + P(G|not L)P(not L)].

3. Estimate parameters

Discuss how to estimate the prior and likelihoods from data: e.g., historical distribution of reviewer types, average probability of good reviews for lazy vs. diligent reviewers. If data is unavailable, propose reasonable values or ranges and perform sensitivity analysis.

4. Compute and interpret

Plug in the numbers to compute the posterior probability. Interpret the result in context: e.g., if the prior is low, even three good reviews may not imply high probability of laziness. Discuss how the result changes with different assumptions.

5. Discuss limitations and extensions

Acknowledge limitations: independence assumption, binary classification of reviewers, and lack of ground truth. Suggest extensions: incorporating more reviews, using a continuous measure of laziness, or applying a hierarchical Bayesian model.

Key Points to Mention

  • Bayes' theorem and conditional probability
  • Prior probability and base rate fallacy
  • Likelihood functions for lazy vs. diligent reviewers
  • Independence assumption of reviews
  • Sensitivity analysis and robustness of assumptions
  • Practical implications for product analytics (e.g., detecting low-quality reviews)

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

Q5

How does the probability that a reviewer is lazy change as the number of consecutive good reviews grows toward infinity?

Product Analytics & MetricsAlgorithms & Data Structures
Author's notes

The ratio of 1 to the power of N versus 0.6 to the power of N blows up as N increases, so the posterior probability of being lazy approaches 1.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Frame the problem as a probabilistic inference about a reviewer's hidden state (lazy vs. diligent) given a sequence of good reviews. Use Bayes' theorem with a prior on laziness and a likelihood model for good reviews, then analyze the limit as the number of consecutive good reviews goes to infinity. Show that under reasonable assumptions, the posterior probability of laziness tends to 0, but discuss edge cases where it may not (e.g., if a lazy reviewer can also produce good reviews with some probability).

Pro tip: Explicitly state your assumptions about the prior and the likelihoods, and note that the result depends on them; this demonstrates rigor and avoids overgeneralizing. Also, connect the analysis to practical implications for product metrics, such as how many reviews are needed to confidently classify a reviewer as diligent.

1. Define the problem and variables

Clarify what 'lazy' means (e.g., a reviewer who always gives good reviews regardless of quality) and define the event of a 'good review'. Let L be the event that a reviewer is lazy, and G_i be the event that the i-th review is good.

2. Set up a Bayesian model

Assume a prior probability P(L) and likelihoods P(G|L) and P(G|not L). For a lazy reviewer, P(G|L) might be high (e.g., 1 if they always give good reviews), while for a diligent reviewer, P(G|not L) is lower (e.g., p < 1).

3. Apply Bayes' theorem for n consecutive good reviews

Compute the posterior P(L | G_1, ..., G_n) using Bayes' rule, assuming conditional independence of reviews given the reviewer type. The posterior is proportional to P(L) * P(G|L)^n.

4. Analyze the limit as n → ∞

Take the limit of the posterior as n goes to infinity. If P(G|L) > P(G|not L), the posterior tends to 1; if P(G|L) < P(G|not L), it tends to 0; if equal, it stays at the prior. Typically, a lazy reviewer is more likely to give good reviews, so P(G|L) > P(G|not L), and the probability of laziness tends to 1. However, if 'lazy' means they give good reviews with lower probability, the opposite holds.

5. Discuss implications and edge cases

Interpret the result: if the probability of laziness tends to 0, it means we become increasingly confident the reviewer is diligent. Mention that the result hinges on the likelihood ratio and that in practice, we may need a finite number of reviews to reach a decision threshold.

Key Points to Mention

  • Bayes' theorem and posterior probability
  • Prior and likelihood assumptions
  • Conditional independence of reviews given reviewer type
  • Limit of posterior as n → ∞
  • Likelihood ratio and its role in determining the limit
  • Practical implications for product analytics (e.g., number of reviews needed to classify reviewers)

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