← Meta Interview Insights

Meta·Data Scientist·Onsite - System Design / Architecture·Senior

SeniorPrefer not to say
Sep 2025Remote

Summary

This was a deep technical screen for a DS role at Meta, centered entirely on one massive open-ended ML system design problem. The question covered basically every angle you can imagine: features, labeling, modeling choices, evaluation, privacy, and deployment. Felt more like a take-home prompt than a 45-minute conversation.

Questions Asked (6)

Q1

How would you design the features for a model that classifies whether a Facebook session is happening at home, in an office, or at a public venue, using only privacy-preserving signals like timestamps, ASN/ISP type, coarse location, device mix, and contact graph patterns?

System DesignData ModelingTechnical Trade-offs
Author's notes

This was the core of the whole interview and it's genuinely hard.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by framing the problem as a multi-class classification task with privacy constraints, then outline a feature engineering pipeline that transforms raw signals into meaningful behavioral patterns. Emphasize the need for robust validation and discuss trade-offs between model complexity and interpretability, given the sensitive nature of the data.

Pro tip: Highlight that privacy-preserving signals are inherently noisy and biased, so you should propose techniques like differential privacy or federated learning to mitigate risks while maintaining utility. Also, mention the importance of aligning with Meta's privacy principles and legal frameworks like GDPR.

1. Define the problem and constraints

Clarify the classification objective, the three classes (home, office, public), and the privacy constraints that limit data usage. Discuss how to handle label ambiguity and potential biases.

2. Feature engineering from privacy-preserving signals

Derive features from timestamps (e.g., time of day, day of week patterns), ASN/ISP type (e.g., residential vs. corporate vs. mobile), coarse location (e.g., density, stability), device mix (e.g., number and types of devices), and contact graph patterns (e.g., co-location frequency, graph centrality).

3. Model selection and training

Choose appropriate models (e.g., gradient boosted trees, neural networks) that can handle heterogeneous features and multi-class output. Consider privacy-preserving training methods like federated learning or differential privacy.

4. Evaluation and validation

Design evaluation metrics (e.g., accuracy, F1, confusion matrix) and validation strategies (e.g., temporal splits, cross-validation) that account for privacy constraints and potential distribution shifts.

5. Deployment and monitoring

Discuss how to deploy the model in production while ensuring privacy, and set up monitoring for performance drift and privacy leakage. Consider interpretability and user consent.

Key Points to Mention

  • Privacy-preserving techniques: differential privacy, federated learning, anonymization
  • Feature engineering: temporal patterns, ASN classification, device fingerprinting, graph-based features
  • Handling class imbalance and label noise
  • Model interpretability and explainability for trust and compliance
  • Evaluation metrics and validation strategies for privacy-sensitive data
  • Ethical and legal considerations (GDPR, CCPA, Meta's privacy principles)

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

Q2

How would you generate labels at scale for this network-context classification problem without ground truth, and how do you correct for bias in those noisy weak labels?

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

Weak supervision is something I'd read about but never actually built end to end.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by outlining a scalable weak supervision pipeline that combines multiple labeling functions (heuristics, distant supervision, and model-based signals) to generate noisy labels efficiently. Then describe a bias correction strategy using probabilistic models like Snorkel or expectation-maximization to estimate and adjust for each source's accuracy and correlation. Emphasize validation with a small gold-standard set and iterative refinement to ensure label quality.

Pro tip: Quantify the trade-off between label quantity and quality: show how you'd measure the marginal value of additional weak labels versus the cost of noise, and propose a feedback loop where model predictions inform label correction. This demonstrates product sense and technical depth.

1. Define labeling sources and coverage

Identify scalable sources such as user behavior logs, metadata, rule-based heuristics, and pre-trained models. Assess their coverage and potential biases relative to the target network-context classification task.

2. Generate weak labels at scale

Apply labeling functions to produce noisy labels for a large dataset. Use a framework like Snorkel to combine multiple signals and output probabilistic labels, avoiding manual annotation bottlenecks.

3. Model and correct label noise

Use a generative model (e.g., label model) to estimate each labeling function's accuracy and correlation, then infer true labels. Alternatively, apply techniques like co-training or noise-aware loss functions during training.

4. Validate and iterate with gold data

Curate a small, high-quality validation set to measure label quality and model performance. Analyze residual bias by comparing weak label distributions to gold labels and adjust labeling functions or model parameters accordingly.

5. Monitor and refine in production

Deploy the model with continuous monitoring for drift and bias. Use active learning to selectively annotate uncertain or biased samples, feeding corrections back into the weak supervision pipeline.

Key Points to Mention

  • Weak supervision techniques (e.g., Snorkel, labeling functions, distant supervision)
  • Probabilistic graphical models for estimating label accuracy and correlation
  • Bias correction methods (e.g., reweighting, adversarial debiasing, noise-aware training)
  • Use of a small gold-standard set for validation and bias detection
  • Scalability considerations: distributed processing, cost of labeling sources
  • Iterative feedback loops and active learning for continuous improvement

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

Q3

Compare rule-based classifiers, gradient boosted trees, and sequence models like HMMs or transformers for this session-context classification task. How do you handle aggregating session-level predictions up to the household level?

System DesignTechnical Trade-offsData Modeling
Author's notes

I went rule lists first as a sanity baseline, then GBTs for the tabular feature regime.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the task and data characteristics (e.g., session length, feature types, label availability). Then compare the three model families in terms of accuracy, interpretability, scalability, and data requirements, and finally discuss aggregation strategies from session to household level, emphasizing the trade-offs between simple pooling and learned aggregation.

Pro tip: Show awareness of production constraints: rule-based models are fast and interpretable but brittle; GBMs handle tabular features well but may miss sequential patterns; transformers are powerful but data-hungry and expensive. For aggregation, mention that household-level labels are often sparse, so consider multi-instance learning or attention-based pooling to weight sessions by relevance.

1. Clarify the problem and data

Ask about session definition, available features (e.g., time, device, content), label granularity (session vs. household), and data volume. This ensures the comparison is grounded in the actual task.

2. Compare model families

Evaluate rule-based classifiers (interpretable, fast, but limited), gradient boosted trees (strong on tabular data, handle non-linearity, but ignore sequence), and sequence models (HMMs for simple sequential dependencies, transformers for long-range context). Discuss trade-offs in accuracy, interpretability, training cost, and data needs.

3. Address session-level prediction

Explain how each model would produce session-level predictions, including feature engineering (e.g., session embeddings for GBMs) and handling of variable-length sequences (e.g., padding/masking for transformers).

4. Aggregate to household level

Propose aggregation methods: simple pooling (mean, max, vote), weighted pooling (by session length or confidence), or learned aggregation (e.g., attention over sessions, hierarchical models). Discuss how to handle households with few sessions and the need for household-level labels.

5. Recommend and justify

Based on the clarified constraints, recommend a model and aggregation strategy, highlighting potential pitfalls (e.g., data leakage, overfitting) and evaluation metrics (e.g., household-level AUC).

Key Points to Mention

  • Trade-offs between interpretability (rule-based) and predictive power (transformers)
  • Handling variable-length sessions and missing data in sequence models
  • Feature engineering for GBMs to capture sequential patterns (e.g., session statistics)
  • Aggregation methods: mean/max pooling, attention-based pooling, multi-instance learning
  • Household-level label scarcity and how to leverage session-level labels
  • Scalability and latency considerations for production deployment

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

Q4

How would you evaluate this model? Define your metrics, describe how you'd set up cross-geo temporal cross-validation, and explain how you'd prevent label leakage from Portal adoption data.

A/B Testing & ExperimentationProduct Analytics & MetricsTechnical Trade-offs
Author's notes

The leakage question from Portal adoption was the sharpest part.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the model's objective and the business context, then propose a metric suite that balances offline predictive performance with online business impact. For validation, design a cross-geo temporal scheme that respects time and geography, and explicitly address label leakage by defining labels and features with strict temporal separation and excluding adoption-derived signals.

Pro tip: Emphasize that offline metrics are proxies; the ultimate test is a well-designed online experiment (A/B test) with guardrail metrics. Also, mention that you would simulate the cross-validation scheme on historical data to verify it prevents leakage before deploying.

1. Clarify Objective and Metrics

Define the model's goal (e.g., predict Portal adoption or a related outcome) and select metrics that align with business impact, such as AUC, precision@k, or lift, alongside online metrics like adoption rate and engagement.

2. Design Cross-Geo Temporal Validation

Split data by both time and geography: train on earlier periods and some geos, validate on later periods and held-out geos. This tests temporal generalization and geographic robustness, mimicking real-world deployment.

3. Prevent Label Leakage

Ensure features are computed only from data available before the label window. Exclude any Portal adoption data that occurs after the prediction time, and avoid using future adoption signals as features.

4. Evaluate and Iterate

Assess model performance using the chosen metrics, compare against baselines, and analyze errors across geos and time. Use insights to refine features and model.

5. Plan Online Validation

Propose an A/B test to measure causal impact on business metrics, with proper randomization and guardrails, to confirm offline findings.

Key Points to Mention

  • Choice of metrics: offline (AUC, precision/recall) and online (adoption rate, engagement, retention)
  • Cross-geo temporal cross-validation: train on past data from some geos, validate on future data from other geos
  • Label leakage prevention: strict temporal separation between features and labels; exclude adoption data from feature set
  • Handling geographic differences: include geo-specific features or use geo as a grouping variable in validation
  • Online experimentation: A/B test with proper power analysis and guardrail metrics
  • Trade-offs: offline metrics may not perfectly correlate with online impact; need to balance model complexity and interpretability

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

Q5

What privacy and compliance constraints would you apply to this system, including data minimization, retention limits, on-device inference options, and re-identification risk?

Technical Trade-offsAdaptability & AmbiguitySystem Design
Author's notes

Probably my weakest section.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the system's purpose and data flows, then systematically address each privacy dimension (minimization, retention, on-device, re-identification) with concrete trade-offs. Tie your recommendations to Meta's regulatory context (GDPR, CCPA) and business goals, showing you can balance innovation with compliance.

Pro tip: Emphasize that privacy constraints are not just legal requirements but can be a competitive advantage—e.g., on-device inference reduces latency and improves user trust, which Meta values. Also, mention that you'd document assumptions and iterate with legal/privacy teams early.

1. Clarify system and data flows

Ask questions to understand what data the system uses, where it comes from, and how it's processed. Map out data collection, storage, and sharing points to identify privacy risks.

2. Apply data minimization and purpose limitation

Propose collecting only data strictly necessary for the model's function, and ensure it's used only for stated purposes. Discuss techniques like feature selection, aggregation, and synthetic data.

3. Define retention limits and deletion policies

Recommend specific retention periods based on legal requirements and business needs, with automated deletion. Mention the need for audit trails and user-initiated deletion.

4. Evaluate on-device inference and edge processing

Assess whether parts of the model can run on-device to avoid sending raw data to servers. Discuss trade-offs in model size, accuracy, and update mechanisms.

5. Mitigate re-identification risks

Propose techniques like k-anonymity, differential privacy, and data masking. Explain how you'd test for re-identification vulnerabilities and monitor for emerging risks.

Key Points to Mention

  • Data minimization: collect only what's needed, use aggregation or synthetic data where possible.
  • Retention limits: set specific timeframes (e.g., 30 days for raw logs), automate deletion, and provide user controls.
  • On-device inference: benefits like reduced latency, improved privacy, and lower server costs; challenges like model compression and update distribution.
  • Re-identification risk: apply differential privacy, k-anonymity, and regular audits; consider auxiliary data that could enable linkage attacks.
  • Compliance frameworks: GDPR, CCPA, and Meta's own privacy principles; mention data protection impact assessments (DPIAs).
  • Trade-offs: balance privacy with model performance and business metrics; propose A/B testing or privacy-preserving analytics.

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

Q6

How would you deploy this system, monitor for drift, and design a holdout experiment to measure whether location-type targeting actually improves Portal ad conversion?

A/B Testing & ExperimentationSystem DesignProduct Strategy
Author's notes

Ended on this and I was running low on steam.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Structure your answer around three pillars: deployment, drift monitoring, and experiment design. For deployment, describe a scalable, low-latency serving architecture with logging; for drift, outline metrics and alerts; for the holdout experiment, define a clean randomized design with proper power analysis and guardrail metrics. Tie everything back to the business goal of improving Portal ad conversion via location-type targeting.

Pro tip: Emphasize that the holdout must be a true control (no targeting) and that you’d pre-register the analysis plan to avoid p-hacking. Also mention that drift monitoring should include both data drift and concept drift, with automated retraining triggers.

1. Deployment Architecture

Describe a real-time serving system (e.g., feature store + model server) that logs predictions, features, and outcomes. Ensure scalability, low latency, and versioning for reproducibility.

2. Drift Monitoring

Define metrics for data drift (e.g., PSI, KL divergence on location-type features) and concept drift (e.g., conversion rate by segment). Set up alerts and a dashboard, with automated retraining when thresholds are breached.

3. Holdout Experiment Design

Randomize users or ad impressions into treatment (location-type targeting) and control (no targeting or baseline). Determine sample size via power analysis, and pre-register primary (conversion rate) and guardrail metrics (e.g., user experience, revenue).

4. Analysis and Decision

Analyze results with appropriate statistical tests (e.g., t-test, CUPED for variance reduction). Check for novelty effects and segment-level heterogeneity. Decide whether to launch based on statistical and practical significance.

Key Points to Mention

  • Use of a feature store to ensure consistent features between training and serving.
  • Monitoring both data drift (input distribution) and concept drift (relationship between features and target).
  • Randomization unit: user-level vs. impression-level, and potential interference.
  • Power analysis to determine sample size and experiment duration.
  • Guardrail metrics to ensure no harm to user experience or other business metrics.
  • Pre-registration of analysis plan to maintain statistical rigor.

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