← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
Apr 2026

Summary

Meta DS interview focused almost entirely on a video recommendation push system, mixing probability theory with product thinking and ML model selection. Pretty math-heavy for what I expected to be more of a product-flavored round.

Questions Asked (4)

Q1

Given an inventory of N videos, what is the probability that a specific ordered sequence of k videos gets pushed to a user? And what changes if the order doesn't matter?

Algorithms & Data StructuresProduct Analytics & Metrics
Author's notes

The ordered vs unordered distinction is where I slipped up a little.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

First, clarify the assumptions: are videos sampled without replacement (each video can appear at most once) or with replacement? Then compute the probability for the ordered case as 1 / P(N, k) if without replacement, or 1 / N^k if with replacement. For the unordered case, multiply by k! to account for all permutations, giving k! / P(N, k) = 1 / C(N, k) without replacement, or k! / N^k with replacement.

Pro tip: Always state your assumptions explicitly and discuss how the answer changes if videos can repeat. Also, relate this to real-world recommendation systems where probabilities are not uniform and order may matter for user engagement.

1. Clarify assumptions

Ask whether videos are sampled with or without replacement, and whether each video is equally likely to be selected. This determines the probability space.

2. Compute ordered probability

For without replacement, the number of ordered sequences of length k from N is P(N, k) = N! / (N-k)!. So the probability of a specific sequence is 1 / P(N, k). For with replacement, it is 1 / N^k.

3. Compute unordered probability

If order doesn't matter, the number of unordered sets of size k is C(N, k) = N! / (k! (N-k)!). The probability of a specific set is 1 / C(N, k) without replacement. With replacement, it is k! / N^k (since each set corresponds to k! ordered sequences).

4. Discuss implications and extensions

Mention that in practice, recommendation systems use non-uniform probabilities and may consider order for ranking. Also, note that if k > N without replacement, probability is zero.

Key Points to Mention

  • Permutations vs combinations: P(N, k) = N! / (N-k)! and C(N, k) = N! / (k! (N-k)!).
  • With replacement: ordered probability = 1 / N^k, unordered probability = k! / N^k.
  • Without replacement: ordered probability = 1 / P(N, k), unordered probability = 1 / C(N, k).
  • Assumption of uniform distribution: each video equally likely to be selected.
  • Real-world relevance: recommendation systems often have non-uniform probabilities and order may affect user engagement.
  • Edge cases: k > N without replacement gives probability 0; k = 0 gives probability 1.

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

Q2

Given the probability of an event, how do you compute its complement, and how does that apply to this video selection scenario?

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

Straightforward complementary probability, 1 minus p.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clearly stating the complement rule: P(not A) = 1 - P(A). Then, apply it to the video selection scenario by defining the event of interest (e.g., a user selects a video) and its complement (e.g., no video selected), and explain how this helps compute probabilities for A/B testing metrics like click-through or selection rates.

Pro tip: Emphasize that the complement rule is especially useful when the event of interest is rare or when direct computation is complex; in A/B testing, it simplifies calculating conversion rates and expected non-conversions, which are key for power analysis.

1. State the complement rule

Clearly articulate that the probability of an event not occurring is 1 minus the probability of the event occurring: P(A^c) = 1 - P(A).

2. Define the event in the scenario

Identify the specific event in the video selection scenario, such as a user selecting a video from a set of options, and denote its probability as p.

3. Compute the complement

Calculate the probability of the complement (e.g., user does not select the video) as 1 - p. This is often easier when p is close to 1 or when non-selection is the focus.

4. Apply to A/B testing metrics

Explain how this complement probability is used in A/B testing, such as computing expected non-conversions, sample size calculations, or analyzing selection rates between control and treatment groups.

5. Discuss practical implications

Highlight why this matters: it helps in interpreting results, ensuring statistical power, and making data-driven decisions about video selection features.

Key Points to Mention

  • Complement rule: P(A^c) = 1 - P(A)
  • Application to video selection: e.g., probability of selecting a video vs. not selecting
  • Use in A/B testing: conversion rates, non-conversion rates, power analysis
  • Independence assumption: if multiple videos, complement may involve joint probabilities
  • Edge cases: when P(A) = 0 or 1, complement is trivial
  • Communication: explain to non-technical stakeholders in simple terms

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

Q3

Should the system push the same video to two friends, or different videos? Walk through the tradeoffs and what metrics you'd expect to move.

Product Sense & IdeationA/B Testing & ExperimentationProduct Analytics & Metrics
Author's notes

This was the part I actually enjoyed.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product context and the goal of the recommendation system, then frame the decision as a tradeoff between short-term engagement and long-term user satisfaction. Walk through the tradeoffs systematically, and propose metrics that capture both immediate and downstream effects, including potential A/B tests to validate.

Pro tip: Emphasize that the 'right' answer depends on the platform's objective (e.g., time spent vs. meaningful social interactions) and that you would run a long-term holdout experiment to measure cumulative impact, not just short-term clicks.

1. Clarify the product and goal

Ask clarifying questions about the platform (e.g., Facebook Watch, Instagram Reels), the recommendation system's objective (e.g., maximize engagement, foster connections), and the definition of 'friends' (close friends vs. all friends).

2. Identify tradeoffs

Discuss the tradeoffs: same video may increase shared experience and social discussion but risks redundancy and reduced content diversity; different videos may increase discovery and individual relevance but could reduce common ground and social interaction.

3. Map to metrics

Propose metrics for each tradeoff: for same video, metrics like video completion rate, shares, comments, and friend interactions; for different videos, metrics like watch time, likes, diversity of content consumed, and user retention.

4. Consider user segments and context

Segment users by relationship strength, content preferences, and usage patterns. For example, close friends might benefit from same video, while casual friends might prefer different videos to avoid redundancy.

5. Design an experiment

Outline an A/B test: control (same video to both friends) vs. treatment (different videos). Define primary and guardrail metrics, and discuss potential long-term effects and network effects.

Key Points to Mention

  • Short-term engagement metrics (CTR, watch time) vs. long-term satisfaction and retention
  • Social interaction metrics (comments, shares, messages between friends)
  • Content diversity and discovery as a driver of long-term engagement
  • Potential for redundancy and annoyance if same video is pushed
  • Network effects and spillover: how one user's behavior affects friends
  • Importance of A/B testing with a long-term holdout to measure cumulative impact

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

Q4

What statistical or ML model would you use for this video recommendation problem, and what's your reasoning?

Technical Trade-offsSystem Design
Author's notes

I went with collaborative filtering as the baseline and then mentioned sequence models for capturing watch history patterns.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the problem scope—what signals are available, what the objective is (e.g., engagement, watch time), and constraints like latency and scale. Then propose a two-stage architecture: candidate generation with a lightweight model (e.g., matrix factorization or two-tower) and ranking with a more expressive model (e.g., gradient boosted trees or deep neural networks). Justify each choice by tying it to the problem characteristics and trade-offs.

Pro tip: Emphasize that the best model depends on the specific business objective and data—show you can adapt rather than defaulting to the fanciest algorithm. Mention how you'd evaluate offline metrics and design online A/B tests to validate improvements.

1. Clarify the problem and constraints

Ask about the recommendation goal (e.g., maximize watch time, clicks), available data (user interactions, video features, context), and system constraints (latency, scale, cold-start). This ensures your model choice is grounded in reality.

2. Propose a two-stage architecture

Explain that at scale, a single model is often impractical. Use a candidate generation stage to narrow millions of videos to hundreds, then a ranking stage to order them. This balances efficiency and accuracy.

3. Select models for each stage with reasoning

For candidate generation, suggest a two-tower neural network or matrix factorization for its efficiency in retrieving relevant items. For ranking, propose a deep neural network (e.g., DLRM) or gradient boosted trees to capture complex feature interactions and optimize the business metric.

4. Discuss trade-offs and alternatives

Acknowledge trade-offs: two-tower models are fast but less expressive; deep models are powerful but require more data and compute. Mention alternatives like session-based RNNs or transformers for sequential patterns, and explain when they might be preferred.

5. Outline evaluation and iteration

Describe offline evaluation using metrics like recall@k, NDCG, or AUC, and online A/B testing to measure impact on business metrics. Highlight the importance of monitoring for bias, freshness, and feedback loops.

Key Points to Mention

  • Two-stage architecture (candidate generation + ranking) for scalability
  • Two-tower neural networks for efficient candidate retrieval
  • Deep learning models (e.g., DLRM) for ranking with feature interactions
  • Handling cold-start with content-based features or meta-learning
  • Offline metrics (recall@k, NDCG) and online A/B testing for evaluation
  • Trade-offs between model complexity, latency, and interpretability

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