Design and Validate Initial Restaurant Recommendation Model

Quick Overview

Evaluates machine learning product judgment for a first restaurant recommendation model on Facebook. Strong answers cover candidate generation, features, labels, cold start, offline validation, online experiments, diversity, privacy, and how to decide when metrics move in opposite directions.

Design and Validate Initial Restaurant Recommendation Model

Company: Meta

Role: Data Scientist

Category: Machine Learning

Difficulty: medium

Interview Round: Onsite

##### Scenario First-iteration machine-learning model for restaurant recommendations on Facebook. ##### Question Describe how you would design an initial recommendation model for restaurants. How would you validate whether the model is working? If one performance metric rises while another drops, how do you interpret and act? ##### Hints Cold-start, features, offline vs online validation, multi-objective optimisation.

Quick Answer: Evaluates machine learning product judgment for a first restaurant recommendation model on Facebook. Strong answers cover candidate generation, features, labels, cold start, offline validation, online experiments, diversity, privacy, and how to decide when metrics move in opposite directions.

|Home/Machine Learning/Meta
Meta logo
Meta
Jul 12, 2025, 6:59 PM
mediumData ScientistOnsiteMachine Learning
28
0

Design and Validate an Initial Restaurant Recommendation Model

You are designing a first-iteration machine-learning model to recommend restaurants to Facebook users on surfaces such as local discovery, Search, or a recommendation module.

Constraints & Assumptions

  • Build a pragmatic first version rather than a perfect recommender system.
  • Assume access to user interactions, restaurant Page data, location context where permitted, social graph signals, and historical restaurant-related actions.
  • The system must balance relevance, diversity, quality, latency, privacy, and fairness across businesses.
  • Validation should include both offline model quality and online product impact.

Clarifying Questions to Ask Guidance

  • What user action is the model optimizing: click, save, call, route, reservation, order, or repeat visit?
  • Where will recommendations be shown, and how many items can be displayed?
  • Are there explicit privacy restrictions on location or social signals?
  • How much training data exists for restaurant interactions in the target market?

Part 1 - Design the Initial Model

Describe how you would design an initial recommendation model for restaurants, including candidate generation, features, labels, and cold-start handling.

What This Part Should Cover Guidance

  • Candidate generation using location, restaurant category, popularity, social signals, and business-quality filters.
  • Features from user preferences, restaurant attributes, context, social graph, freshness, distance, price, and quality.
  • Label choices tied to user value, such as saves, clicks with dwell, calls, directions, bookings, or repeat actions.
  • Cold-start strategies for new users, new restaurants, and sparse markets.

Part 2 - Validate the Model

How would you validate whether the model is working?

What This Part Should Cover Guidance

  • Offline evaluation with train/test splits, ranking metrics, calibration, segment checks, and leakage prevention.
  • Online experiments measuring incremental user actions, satisfaction, retention, and downstream business outcomes.
  • Guardrails for hides, irrelevant recommendations, local business quality, latency, diversity, and privacy-sensitive features.

Part 3 - Interpret Metric Trade-offs

If one performance metric rises while another drops, how would you interpret the trade-off and decide what to do?

What This Part Should Cover Guidance

  • Examples such as click-through rate rising while saves, repeat usage, diversity, or complaint rate worsens.
  • Decision rules based on primary metric, guardrail thresholds, segment-level impact, and long-term value.
  • Follow-up analyses to identify whether the model is exploiting shallow engagement instead of true relevance.

What a Strong Answer Covers Guidance

A strong answer shows a complete recommender workflow: candidates, features, labels, cold start, offline validation, online experimentation, guardrails, and principled trade-off decisions.

Follow-up Questions Guidance

  • How would you prevent popular restaurants from dominating all recommendations?
  • How would you detect if the model is biased toward certain neighborhoods or business types?
  • What would be your simplest launchable baseline model?
Loading comments...