← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Intermediate

Intermediate
Apr 2026

Summary

Analytical execution round for a Data Scientist role at Meta's ads team, heavy on probability theory and statistics. The whole thing felt like a grad school exam more than a job interview, which I was not fully prepared for.

Questions Asked (3)

Q1

Name three probability distributions suitable for modeling ad clicks, state the core assumption behind each, and write out the expectation and variance formulas for all three.

Product Analytics & MetricsA/B Testing & Experimentation
Author's notes

I went with Bernoulli, Binomial, and Poisson, which felt right.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by naming three distributions: Bernoulli, Binomial, and Poisson. For each, clearly state the core assumption (e.g., single trial, fixed number of independent trials, or rare events in a fixed interval) and then provide the expectation and variance formulas. Emphasize how these distributions relate to ad click modeling and when each is appropriate.

Pro tip: Mention that in practice, ad clicks often exhibit overdispersion (variance > mean), so the Poisson assumption may be too restrictive; consider negative binomial or quasi-Poisson as extensions. This shows awareness of real-world data complexities.

1. Identify distributions

Select three distributions: Bernoulli, Binomial, and Poisson. These cover single click, clicks out of impressions, and clicks per time period.

2. State core assumptions

For each, articulate the key assumption: Bernoulli: one trial with constant click probability; Binomial: fixed number of independent trials with constant probability; Poisson: events occur independently at a constant rate in a fixed interval.

3. Provide expectation and variance

Write the formulas: Bernoulli: E[X]=p, Var(X)=p(1-p); Binomial: E[X]=np, Var(X)=np(1-p); Poisson: E[X]=λ, Var(X)=λ.

4. Relate to ad clicks

Explain how each models ad clicks: Bernoulli for a single impression, Binomial for multiple impressions with same click probability, Poisson for clicks over time or space when events are rare.

5. Discuss limitations and extensions

Note that real ad click data may violate assumptions (e.g., overdispersion, varying probabilities), and mention alternatives like negative binomial or beta-binomial.

Key Points to Mention

  • Bernoulli distribution models a single ad impression with click probability p.
  • Binomial distribution models the number of clicks in n independent impressions with constant click probability p.
  • Poisson distribution models the number of clicks in a fixed time interval when clicks occur independently at a constant rate λ.
  • Expectation and variance formulas: Bernoulli (p, p(1-p)), Binomial (np, np(1-p)), Poisson (λ, λ).
  • Assumptions: independence, constant probability/rate, fixed number of trials or interval.
  • Overdispersion in real data and potential use of negative binomial or quasi-Poisson.

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

Q2

If 90% of users are high-intent and 10% are low-intent, and both groups have an expected click-through rate of 0.30, what is the overall expected CTR?

Product Analytics & MetricsA/B Testing & Experimentation
Author's notes

Straightforward law of total expectation application.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Recognize that the overall expected CTR is a weighted average of the group CTRs, where the weights are the group proportions. Since both groups have the same CTR (0.30), the overall expected CTR is simply 0.30. State the calculation clearly and note that the result is independent of the mix when group CTRs are equal.

Pro tip: Even when the answer seems trivial, briefly mention that if the CTRs differed, you would compute a weighted average, and that this logic extends to any segmentation. This shows you understand the general principle and can handle more complex scenarios.

1. Identify the given information

Restate the proportions (90% high-intent, 10% low-intent) and the CTR for each group (0.30).

2. Recall the formula for overall CTR

The overall expected CTR is the weighted average: (proportion_high * CTR_high) + (proportion_low * CTR_low).

3. Plug in the numbers

Compute (0.9 * 0.30) + (0.1 * 0.30) = 0.27 + 0.03 = 0.30.

4. Interpret the result

Explain that because both groups have the same CTR, the overall CTR equals that common value, regardless of the mix.

5. Generalize (optional)

Mention that if the CTRs were different, the overall CTR would be a weighted average that depends on the proportions.

Key Points to Mention

  • Weighted average formula for overall CTR
  • The calculation: 0.9*0.30 + 0.1*0.30 = 0.30
  • The result is independent of the mix when group CTRs are equal
  • Extension to unequal CTRs: overall CTR would be between the two group CTRs, weighted by proportions
  • Importance of segment-level metrics in product analytics
  • Clear communication of the trivial case without overcomplicating

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

Q3

Sketch the PDF of an exponential distribution and explain what happens to its shape as sample size grows very large.

A/B Testing & ExperimentationProduct Analytics & Metrics
Author's notes

The sketch part was fine, decreasing curve starting at lambda, everyone knows that.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by sketching the exponential PDF with its characteristic decreasing curve and long right tail, clearly labeling the rate parameter λ and the mean 1/λ. Then explain that as sample size grows, the distribution of the sample mean (not the raw data) approaches a normal distribution due to the Central Limit Theorem, and relate this to why A/B tests rely on normal approximations for large samples.

Pro tip: Emphasize that the exponential distribution itself does not change shape with sample size—it's the sampling distribution of the mean that becomes normal. This distinction shows you understand the difference between the population distribution and the sampling distribution, which is crucial for experiment analysis.

1. Sketch the Exponential PDF

Draw a curve starting at λ on the y-axis when x=0, decaying exponentially toward zero as x increases, with a long right tail. Label the x-axis as the variable (e.g., time between events) and the y-axis as density, and mark the mean at 1/λ.

2. Explain the Effect of Sample Size

Clarify that the shape of the exponential distribution itself remains unchanged regardless of sample size. Instead, as sample size grows, the distribution of the sample mean becomes approximately normal with mean 1/λ and variance 1/(nλ²), by the Central Limit Theorem.

3. Connect to A/B Testing

Discuss how this justifies using normal-based methods (e.g., t-tests, z-tests) for large samples in A/B tests, even when the underlying metric (like time between clicks) is exponential. Mention that the skewness diminishes in the sampling distribution as n increases.

4. Address Practical Implications

Note that for small samples, the exponential distribution's skewness can bias results, so transformations or non-parametric methods might be needed. For large samples, the CLT ensures robustness, but be aware of heavy tails or extreme outliers.

Key Points to Mention

  • Exponential PDF formula: f(x) = λe^{-λx} for x ≥ 0, with mean 1/λ and variance 1/λ².
  • The exponential distribution is skewed right and has a long tail, which affects small-sample inference.
  • Central Limit Theorem: the sampling distribution of the mean approaches normality as n increases, regardless of the population distribution.
  • The standard error of the mean decreases with √n, improving precision in A/B tests.
  • In practice, A/B tests at Meta often rely on large samples, so normal approximations are valid even for skewed metrics.
  • Be cautious with metrics that have extreme skewness or heavy tails; consider using trimmed means or other robust methods.

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