ML-Powered Feed Features for Facebook: Choosing Users, Signals and Success Metrics
Company: Meta
Role: Product Manager
Category: Product Design & Strategy
Difficulty: medium
Interview Round: Other
You are a product manager on Facebook's Feed (formerly News Feed), the ranked stream of posts people see when they open the app. Machine learning already decides much of what appears there. How would you use ML to make the Feed better?
Choose whom you are improving it for and which of their problems you are solving, describe the ML-powered changes you would make, and explain how you would know they worked.
```hint Start from a frustration, not a model
Name a specific frustration a group of Feed users has. Then ask what the system would need to predict well for that frustration to go away.
```
### Clarifying Questions
- Is the goal more engagement, more value per visit, or bringing back people who are drifting away?
- Should the ideas change how the existing Feed is ranked, or can they be new Feed features or surfaces that use ML?
- Are ads in Feed in scope, or only organic posts from friends, groups, Pages and recommended accounts?
- Which personal signals may the models use, for example messaging activity or life events a user has posted?
### What a Strong Answer Covers
- A goal and a user segment chosen with reasons, and concrete Feed pain points for that segment
- Each ML idea framed as a specific prediction problem, with the signals, labels and feedback it depends on
- Prioritization by impact, confidence and effort, separating ranking changes from new surfaces
- Success metrics that look past raw engagement, with guardrails for content integrity, well-being, privacy and ads revenue
- ML-specific risks such as engagement bait, feedback loops, cold start and sensitive inferences, and how to test safely
### Follow-up Questions
- Your change increases time spent, but survey answers about whether posts were worth seeing get worse. What do you do?
- Your training data only contains posts the old ranking chose to show. How do you evaluate the new model before an online test?
- How would you explain to a user why a post appeared, and how much control over ranking would you give them?
Overview: A product sense question about applying machine learning to the Facebook Feed. It tests whether you can choose a user segment and a real pain point, frame ML ideas as concrete prediction problems with signals and labels, prioritize them, and define success metrics and guardrails beyond raw engagement.