← Meta Interview Insights

Meta·Data Scientist·Technical Phone Screen·Senior

Senior
Jul 2026

Summary

Meta DS interview focused entirely on designing an ad ranking system end-to-end. It was a deep technical session covering modeling, evaluation, and all the messy real-world constraints that come with running ads at scale. Dense question, lots of ground to cover.

Questions Asked (7)

Q1

How would you define the objectives for an ad recommendation system in a content feed, and how would you balance revenue against user experience?

Product StrategyPricing & MonetizationProduct Analytics & Metrics
Author's notes

I started with the obvious revenue angle and then pivoted to user experience, but I fumbled the part about how to actually combine them.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the objectives around a north-star metric that captures long-term user value, then decompose it into revenue and user experience components. Discuss how to balance them through experimentation, guardrail metrics, and multi-objective optimization, emphasizing trade-offs and long-term sustainability.

Pro tip: Show that you understand the two-sided nature of the marketplace: advertisers need ROI, users need relevance, and the platform needs sustainable growth. Mention that short-term revenue gains can harm long-term user retention, so you'd use holdout experiments to measure long-term effects.

1. Define the North Star and Objectives

Identify a north-star metric like long-term user value (LTV) or daily active users (DAU) that aligns with the platform's mission. Break it down into revenue objectives (e.g., ad revenue per user) and user experience objectives (e.g., user satisfaction, engagement).

2. Select Metrics for Each Objective

Choose specific metrics: for revenue, consider CPM, CTR, conversion rate; for user experience, consider engagement metrics (time spent, likes, shares), satisfaction scores, and retention. Ensure metrics are measurable and actionable.

3. Balance via Multi-Objective Optimization

Propose a weighted objective function or constraints to balance revenue and user experience. Use techniques like Pareto optimization or setting guardrail metrics (e.g., user satisfaction must not drop below X).

4. Experiment and Iterate

Design A/B tests to measure the impact of different ad loads or targeting strategies on both revenue and user experience. Use long-term holdout groups to detect delayed effects.

5. Monitor and Adjust

Continuously monitor key metrics and adjust the balance based on business goals and user feedback. Establish a governance process for making trade-off decisions.

Key Points to Mention

  • North-star metric: long-term user value (LTV) or DAU
  • Revenue metrics: CPM, CTR, conversion rate, ad load
  • User experience metrics: engagement, satisfaction, retention, churn
  • Multi-objective optimization and guardrail metrics
  • A/B testing and long-term holdout experiments
  • Trade-offs: short-term revenue vs. long-term user trust

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

Q2

What modeling approach would you use to predict CTR, CVR, and revenue for ads, and how would you handle delayed conversion feedback?

Data ModelingTechnical Trade-offsSystem Design
Author's notes

This is where I felt most comfortable.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by outlining a multi-task learning framework that jointly predicts CTR, CVR, and revenue, leveraging shared representations and task-specific heads. Then discuss how to handle delayed conversion feedback using techniques like survival analysis, importance weighting, or delayed feedback models, emphasizing the trade-offs between bias and variance.

Pro tip: Emphasize the business impact of accurate delayed conversion modeling, such as improved auction dynamics and advertiser ROI, and mention how you would validate the approach using online experiments with long-term holdouts.

1. Define the prediction tasks and data

Clarify that CTR, CVR, and revenue are distinct but related tasks. Identify data sources: user features, ad features, context, and historical conversion delays.

2. Choose a modeling architecture

Propose a multi-task learning model (e.g., shared bottom layers with task-specific heads) or a sequential model where CVR depends on CTR. Discuss trade-offs between joint and separate models.

3. Address delayed conversion feedback

Explain methods like survival analysis (e.g., Cox model), delayed feedback models (e.g., exponential decay), or importance weighting to correct for bias. Mention the need to handle censored data.

4. Evaluate and iterate

Describe offline evaluation metrics (e.g., AUC, calibration, revenue lift) and online validation via A/B tests with long-term holdouts. Discuss how to monitor and update the model as delays change.

Key Points to Mention

  • Multi-task learning to share information across CTR, CVR, and revenue
  • Handling delayed feedback with survival analysis or delayed feedback models
  • Importance weighting to correct for bias from early feedback
  • Trade-offs between model complexity and interpretability
  • Online evaluation with long-term holdouts to capture delayed effects
  • Business impact: improved auction efficiency and advertiser ROI

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

Q3

How would you address selection bias when evaluating a new ranking policy if your training data came from a previous policy's logged decisions?

A/B Testing & ExperimentationTechnical Trade-offs
Author's notes

Counterfactual evaluation.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by acknowledging that selection bias is inherent when training on logged data from a previous policy, then propose a combination of causal inference techniques and experimental design to evaluate the new policy. Emphasize the importance of using counterfactual or off-policy evaluation methods to estimate the new policy's performance without bias.

Pro tip: Mention that you would validate your approach using a small-scale randomized experiment (e.g., an A/B test) to calibrate and correct any residual bias, showing you understand both theory and practical deployment.

1. Identify and Characterize the Bias

Explain how the previous policy's logging creates selection bias, e.g., by only showing items it favored. Discuss the implications for training and evaluation.

2. Choose Appropriate Off-Policy Evaluation Methods

Propose methods like inverse propensity scoring (IPS), doubly robust estimation, or counterfactual modeling to adjust for the bias in logged data.

3. Leverage Experimental Data for Validation

Suggest running a small randomized experiment (A/B test) to collect unbiased data and validate the off-policy estimates, correcting any discrepancies.

4. Iterate and Monitor

Describe how you would continuously monitor the new policy's performance and refine the bias correction as more data becomes available.

Key Points to Mention

  • Selection bias due to logging policy
  • Inverse propensity scoring (IPS)
  • Doubly robust estimation
  • Counterfactual evaluation
  • A/B testing for validation
  • Exploration vs. exploitation trade-off

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

Q4

How would you approach exploration vs. exploitation in an ad ranking system?

Technical Trade-offsProduct Strategy
Author's notes

Went with epsilon-greedy as the starting point and then mentioned contextual bandits.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing exploration vs. exploitation as a fundamental trade-off in ad ranking, where exploration gathers data to improve future rankings while exploitation maximizes immediate relevance and revenue. Then propose a concrete approach, such as using a multi-armed bandit or reinforcement learning framework, and discuss how to balance the two via algorithms like Thompson Sampling or epsilon-greedy, with metrics to evaluate success.

Pro tip: Emphasize that exploration should be deliberate and cost-aware: quantify the expected long-term value of information gained versus short-term revenue loss, and suggest starting with a small exploration budget that scales with confidence.

1. Define the objective and metrics

Clarify that the goal is to maximize long-term user engagement and advertiser value, not just immediate clicks. Define metrics like CTR, conversion rate, and long-term revenue, and acknowledge that exploration may temporarily reduce short-term metrics.

2. Choose an exploration strategy

Select a method such as epsilon-greedy, Thompson Sampling, or upper confidence bounds (UCB) based on the scale and dynamics of the ad system. Explain how each balances exploration and exploitation differently.

3. Design the experimentation framework

Propose an online learning setup where a small fraction of traffic is allocated to exploration, and use techniques like contextual bandits to personalize exploration. Ensure proper logging and feedback loops.

4. Evaluate and iterate

Set up A/B tests or counterfactual evaluations to measure the impact of exploration on long-term metrics. Use off-policy evaluation to safely test new strategies before full deployment.

5. Address practical constraints

Discuss how to handle cold-start, ad fatigue, and budget constraints. Suggest adaptive exploration rates that decrease as confidence in ad performance increases.

Key Points to Mention

  • Multi-armed bandit algorithms (e.g., Thompson Sampling, UCB) and their applicability to ad ranking
  • Contextual bandits to incorporate user and ad features for personalized exploration
  • Trade-off between short-term revenue and long-term learning; use of discount factors or long-term value metrics
  • Off-policy evaluation and counterfactual logging to safely assess new policies
  • Exploration budget allocation and adaptive exploration rates (e.g., decaying epsilon)
  • Cold-start problem for new ads and how exploration helps gather initial data

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

Q5

How would you handle cold start for new ads and new users in your ranking system?

Data ModelingTechnical Trade-offsAdaptability & Ambiguity
Author's notes

For new ads I talked about using advertiser-level features and creative embeddings as proxies until you accumulate clicks.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the two distinct cold start problems: new ads (items) and new users. Then propose a multi-pronged strategy that leverages exploration, content-based features, and transfer learning, while balancing short-term metrics with long-term ecosystem health.

Pro tip: Emphasize that cold start is not just a modeling problem but also a product and experimentation challenge—suggest logging and counterfactual evaluation to measure the impact of exploration without harming user experience.

1. Clarify and scope the problem

Ask clarifying questions to understand the specific ranking system, business goals, and constraints (e.g., latency, fairness, budget). Distinguish between new ads and new users, as their cold start challenges differ.

2. Leverage content and side information

For new ads, use content features (text, image, category) and advertiser history; for new users, use demographics, context, and onboarding signals. This enables content-based ranking when interaction data is sparse.

3. Apply exploration and transfer learning

Use multi-armed bandits or Thompson sampling to explore new ads/users while exploiting known ones. Transfer learning from similar ads/users or meta-learning can provide a warm start.

4. Design evaluation and guardrails

Set up online experiments (A/B tests) with guardrail metrics (e.g., user satisfaction, long-term value) and offline counterfactual evaluation. Monitor for feedback loops and popularity bias.

5. Iterate and scale

Start with a simple heuristic or model, then iterate as data accumulates. Consider hybrid approaches that blend content-based and collaborative signals, and scale across the platform.

Key Points to Mention

  • Exploration-exploitation trade-off (e.g., epsilon-greedy, Thompson sampling, UCB)
  • Content-based filtering and feature engineering for new entities
  • Transfer learning and meta-learning to leverage data from similar ads/users
  • Cold start as a system-level problem: logging, feedback loops, and evaluation
  • Business metrics: short-term CTR vs. long-term user engagement and advertiser value
  • Practical constraints: latency, scalability, and fairness in ranking

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

Q6

How would you incorporate advertiser budget pacing constraints into the ranking and serving system?

System DesignPricing & MonetizationTechnical Trade-offs
Author's notes

Did not have a clean answer here.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a constrained optimization: maximize total advertiser value (e.g., conversions) subject to budget pacing constraints over a time horizon. Then describe how to integrate pacing signals into the ranking function, such as by adjusting bids with a pacing multiplier, and discuss trade-offs between short-term performance and long-term budget delivery.

Pro tip: Emphasize that pacing should be probabilistic and adaptive, not deterministic, to handle uncertainty in traffic and competition. Also, mention the importance of monitoring and feedback loops to avoid overspending or underspending.

1. Define the objective and constraints

Clarify that the goal is to maximize advertiser value (e.g., conversions, clicks) while respecting budget constraints over a specified period (daily, weekly, etc.). Consider both hard constraints (budget limits) and soft constraints (smooth delivery).

2. Model pacing as a control problem

Treat pacing as a feedback control system where the pacing rate adjusts based on observed spend and remaining budget. Use techniques like PID controllers or reinforcement learning to dynamically set pacing multipliers.

3. Integrate pacing into ranking

Incorporate pacing by modifying the bid or score used in ranking. For example, multiply the predicted value by a pacing multiplier that reflects the advertiser's budget status. This ensures that ads with urgent budget needs are prioritized appropriately.

4. Handle trade-offs and system design

Discuss trade-offs: pacing may reduce short-term revenue but improves long-term advertiser satisfaction. Design the system to be scalable and low-latency, possibly using distributed counters and approximate algorithms.

5. Evaluate and iterate

Propose metrics to evaluate pacing effectiveness (e.g., budget utilization, ROI, advertiser retention) and describe A/B testing or simulation to tune parameters. Highlight the need for continuous monitoring and adaptation.

Key Points to Mention

  • Budget pacing as a constrained optimization problem
  • Pacing multipliers or throttling in the ranking function
  • Feedback control mechanisms (e.g., PID, reinforcement learning)
  • Trade-offs between short-term revenue and long-term advertiser value
  • Scalability and latency considerations in a large-scale system
  • Evaluation metrics and A/B testing for pacing strategies

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

Q7

Walk me through your offline and online evaluation plan for this ad ranking system, including what guardrail metrics you would set.

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

Talked about AUC and calibration offline, then revenue per impression and click-through rate as primary online metrics.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by outlining a layered evaluation strategy: offline metrics for model quality, online A/B tests for causal impact, and guardrails to prevent regressions. Emphasize alignment with business goals and user experience, and describe how you would iterate based on results.

Pro tip: Always define guardrail metrics upfront and set clear thresholds for stopping or rolling back experiments; this shows you prioritize long-term ecosystem health over short-term gains.

1. Define Objectives and Metrics

Clarify the primary goal (e.g., increase ad revenue) and select proxy metrics (e.g., CTR, CVR) and guardrail metrics (e.g., user satisfaction, page load time).

2. Offline Evaluation

Use historical data to train and validate the ranking model, evaluating with metrics like AUC, NDCG, and calibration; simulate counterfactuals to estimate online impact.

3. Online Evaluation via A/B Testing

Design a randomized controlled experiment with proper power analysis, randomize at user level, and measure treatment effects on primary and guardrail metrics.

4. Guardrail Monitoring and Decision Rules

Set thresholds for guardrail metrics (e.g., no more than 1% drop in user engagement) and pre-register decision rules for rollback or iteration.

5. Iterate and Scale

Analyze results, diagnose issues, and iterate on the model; if successful, gradually roll out to larger populations while continuing to monitor guardrails.

Key Points to Mention

  • Offline metrics: AUC, NDCG, precision@k, calibration, and counterfactual evaluation (e.g., inverse propensity scoring).
  • Online metrics: CTR, CVR, revenue per user, and ad load; ensure proper randomization and sample size.
  • Guardrail metrics: user engagement (DAU/MAU), session length, page load time, user satisfaction (surveys), and advertiser value.
  • Statistical considerations: power analysis, multiple testing correction, sequential testing, and novelty effects.
  • Long-term holdout groups to measure long-term effects and avoid short-term bias.
  • Cross-functional collaboration with product, engineering, and policy teams to align on metrics and thresholds.

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