← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
May 2026

Summary

Interviewed for a Data Scientist role at Meta and got a stats-heavy question about distribution modeling. Pretty technical, felt like they wanted to see how deep you could go on the theory side, not just name-drop distributions.

Questions Asked (1)

Q1

If you needed to model the number of unique recipients each caller contacts, which statistical distribution would you choose and how would you validate that it fits the data?

Data ModelingTechnical Trade-offsProduct Analytics & Metrics
Author's notes

I went straight to Poisson because that's the reflex answer for count data, but they pushed back pretty fast.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the context: unique recipients per caller is a count of distinct contacts, likely over a fixed time window. Then propose a distribution like Negative Binomial or Poisson, justifying based on overdispersion and the nature of contact behavior. Finally, outline a validation plan using goodness-of-fit tests, visual diagnostics, and out-of-sample checks.

Pro tip: Mention that in practice, you'd also consider zero-inflation or mixture models if many callers have zero or few recipients, and that validating with business metrics (e.g., prediction error on holdout) matters more than pure statistical fit.

1. Clarify the metric and context

Define what 'unique recipients' means (e.g., distinct phone numbers contacted) and the time window (daily, weekly). Consider if the count is per caller and whether there are structural zeros (callers who contact no one).

2. Choose a candidate distribution

Propose Poisson if mean≈variance, but likely Negative Binomial due to overdispersion from heterogeneity in caller behavior. Mention alternatives like Zero-Inflated Poisson/Negative Binomial if many zeros.

3. Validate the fit statistically

Use goodness-of-fit tests (chi-square, Kolmogorov-Smirnov), compare AIC/BIC, and check dispersion statistics. Plot observed vs. expected frequencies and QQ plots.

4. Validate practically and iterate

Split data into train/test, fit on train, and evaluate predictive performance on test (e.g., log-likelihood, MAE). Check if the model captures key business patterns (e.g., heavy tail).

Key Points to Mention

  • Poisson distribution assumes mean equals variance, often violated in real-world count data.
  • Negative Binomial handles overdispersion and is a common choice for count data with heterogeneity.
  • Zero-inflated models if there's an excess of callers with zero unique recipients.
  • Goodness-of-fit tests: chi-square, KS test, and dispersion index.
  • Visual diagnostics: histogram, QQ plot, and rootogram.
  • Cross-validation and out-of-sample prediction to ensure practical utility.

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