Evaluate Factors Before Replacing Recommendation Model

Quick Overview

Evaluates whether an ads recommendation model should replace an existing production model. Strong answers combine offline validation, online experiments, revenue and advertiser outcomes, user guardrails, conflicting metric interpretation, rollout safety, rollback planning, and executive communication.

Evaluate Factors Before Replacing Recommendation Model

Company: Meta

Role: Data Scientist

Category: Machine Learning

Difficulty: hard

Interview Round: Onsite

##### Scenario Facebook Ads team built a new recommendation model and plans to deprecate the old one. ##### Question a) What factors must be evaluated before fully replacing the existing model? b) CTR falls but revenue or other metrics increase—how would you decide which model to ship? c) Outline the A/B-testing procedure you would follow. d) How would you visualise and present the final results to the CFO? ##### Hints Cover offline evaluation, guardrails, rollback plans, stakeholder dashboards, statistical significance.

Quick Answer: Evaluates whether an ads recommendation model should replace an existing production model. Strong answers combine offline validation, online experiments, revenue and advertiser outcomes, user guardrails, conflicting metric interpretation, rollout safety, rollback planning, and executive communication.

|Home/Machine Learning/Meta
Meta logo
Meta
Jul 12, 2025, 6:59 PM
hardData ScientistOnsiteMachine Learning
63
0

Evaluate Factors Before Replacing a Recommendation Model

A large ads platform has built a new recommendation or ranking model and plans to deprecate the existing production model. You need to evaluate whether the replacement is safe and valuable.

Constraints & Assumptions

  • The model affects users, advertisers, platform revenue, and policy or quality outcomes.
  • Assume offline model metrics, online experiment infrastructure, ads delivery logs, user engagement, advertiser conversion data, and revenue metrics are available.
  • A model replacement should not be decided from one metric alone.
  • Include launch, rollback, and executive communication considerations.

Clarifying Questions to Ask Guidance

  • What does the ads model optimize: CTR, conversion value, revenue, advertiser ROI, user utility, or a combined objective?
  • Is the new model a ranking model, candidate generator, calibration model, or targeting model?
  • What are the hard guardrails around user experience, advertiser fairness, policy, and latency?
  • Are there known risks from changing auction dynamics or advertiser budgets?

Part 1 - Evaluate the Replacement

What factors must be evaluated before fully replacing the existing model?

What This Part Should Cover Guidance

  • Offline metrics, calibration, ranking quality, bias, robustness, latency, coverage, freshness, interpretability, and failure modes.
  • Online metrics for revenue, CTR, conversion rate, advertiser ROI, user engagement, complaints, hide/report rate, policy violations, and long-term value.
  • Segment-level impact across advertisers, users, placements, markets, and budget sizes.
  • Operational readiness: monitoring, rollback, logging, and model retraining.

Part 2 - Interpret Conflicting Metrics

If CTR falls but revenue or other business metrics increase, how would you decide which model to ship?

What This Part Should Cover Guidance

  • Predefined objective hierarchy and guardrail thresholds.
  • Understanding whether CTR decline reflects worse user experience or better targeting toward higher-value conversions.
  • Advertiser ROI, user satisfaction, auction health, and long-term revenue effects.
  • Segment-level and long-term analysis before launch.

Part 3 - Design the A/B Test

Outline the A/B testing procedure from canary to full rollout.

What This Part Should Cover Guidance

  • Offline validation, shadow testing, canary rollout, ramp schedule, randomization, sample size, duration, and monitoring.
  • Sample-ratio mismatch checks, instrumentation validation, guardrail alerts, and rollback criteria.
  • Post-launch monitoring for drift and delayed advertiser outcomes.

Part 4 - Present to the CFO

How would you visualize and present the final results to the CFO?

What This Part Should Cover Guidance

  • Clear executive summary of incremental revenue, confidence intervals, downside risks, and guardrail status.
  • Charts showing revenue, advertiser ROI, user guardrails, segment impact, rollout plan, and sensitivity analysis.
  • Recommendation with launch conditions and monitoring commitments.

What a Strong Answer Covers Guidance

A strong answer balances user, advertiser, and platform goals, uses offline and online evidence, handles conflicting metrics through a predefined decision framework, and communicates business impact clearly to executives.

Follow-up Questions Guidance

  • What would make you stop a model ramp immediately?
  • How would you evaluate delayed advertiser conversions?
  • How would you detect that small advertisers are harmed while aggregate revenue improves?
Loading comments...