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

Senior

Jun 2026

Summary

Meta DS interview for a ranking/recommendation role. One massive system design question that covers basically everything you'd ever need to know about ML-powered ranking. It was a lot to hold in your head at once.

Questions Asked (1)

You own the restaurant recommendation surface for a city app. Design a ranking system that balances distance, predicted satisfaction, exploration of new restaurants, and diversity across cuisines. Cover: feature engineering from sparse signals, a learning-to-rank approach that avoids popularity bias and handles cold-start, a feedback-loop mitigation plan, online serving constraints with a fallback, offline/online evaluation with position-bias correction, and how you'd detect and fix geographic and price-point unfairness without hurting user utility.

System DesignProduct Analytics & MetricsTechnical Trade-offs

Author's notes

This question is basically five questions stapled together.

Create a free account to read the full note

AI HintsAI Generated

Suggested Approach

Start by clarifying the product goals and constraints, then propose a multi-objective ranking system that combines learning-to-rank with explicit diversity and exploration mechanisms. Walk through the full lifecycle: feature engineering, model training with bias mitigation, online serving with fallbacks, and rigorous evaluation including fairness audits.

Pro tip: Emphasize that fairness and diversity are not afterthoughts but core ranking objectives; propose a multi-objective optimization framework where these are tuned as hyperparameters to balance user utility and long-term ecosystem health.

1. Clarify Goals and Constraints

Ask clarifying questions about business objectives, user segments, and technical constraints (e.g., latency, data availability). Define success metrics for utility, diversity, and fairness.

2. Feature Engineering and Data Signals

Identify sparse signals (user reviews, check-ins, cuisine tags) and engineer features like distance, predicted satisfaction, cuisine diversity, and novelty. Address sparsity with embeddings and side information.

3. Learning-to-Rank Model with Bias Mitigation

Propose a LTR model (e.g., LambdaMART or neural) that incorporates diversity and exploration via multi-objective loss or re-ranking. Mitigate popularity bias with inverse propensity weighting and handle cold-start via content-based features.

4. Online Serving and Feedback Loops

Design a serving architecture with caching, fallback to heuristic ranking, and exploration (e.g., epsilon-greedy). Implement a feedback-loop mitigation plan with randomization and debiasing.

5. Evaluation and Fairness Auditing

Set up offline evaluation with counterfactual estimators and online A/B tests with position-bias correction. Continuously monitor geographic and price-point fairness, and apply constrained optimization to fix disparities.

Key Points to Mention

Multi-objective ranking: combine relevance, distance, diversity, and exploration into a single utility function with tunable weights.
Cold-start and sparsity: use content-based embeddings, side information, and meta-learning to predict satisfaction for new restaurants/users.
Popularity bias mitigation: apply inverse propensity weighting (IPW) or causal inference to debias training data and avoid rich-get-richer dynamics.
Feedback loop mitigation: introduce randomization in serving, use counterfactual logging, and periodically retrain with debiased data.
Online serving constraints: precompute candidate sets, use approximate nearest neighbor search, and have a fallback heuristic (e.g., distance-based) for low-latency or failure scenarios.
Fairness auditing: define fairness metrics across geographic and price-point segments, monitor via dashboards, and use constrained optimization to reduce disparities while maintaining overall utility.

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