Inferring Co-Viewing on YouTube: When Several People Watch One Screen
Company: Google
Role: Product Manager
Category: Machine Learning
Difficulty: medium
Interview Round: Other
YouTube wants to know whether a video is being watched by one person or by a group, meaning two or more people watching the same playback together. How would YouTube detect group viewing?
Explain which signals you would rely on, how you would turn them into a judgment or an estimate, and how you would check that the detection is accurate.
```hint Start from the use
Whether the result feeds ad measurement, recommendations or a feature viewers see changes how accurate it must be, whether an aggregate estimate is enough and which privacy limits apply. Pin that down before choosing signals.
```
### Clarifying Questions
- What will the detection be used for: measuring how many people an ad reached, tuning recommendations, or a feature viewers see?
- Is an aggregate estimate enough, such as the share of TV watch time that is group viewing, or does each playback need its own label?
- Which devices are in scope: TVs and streaming devices only, or phones and computers too?
- Can YouTube ask viewers directly, and are camera or microphone sensing off the table for privacy reasons?
### What a Strong Answer Covers
- Why YouTube wants the signal, and how the use case sets the bar for accuracy, granularity and privacy.
- A precise definition of group viewing and a credible source of ground-truth labels.
- Observable signals across device, household, time, content and interaction, combined in a calibrated model or estimate.
- Evaluation against labeled data, checks for confounding and bias across segments, and how the output is served and monitored.
- Consent, transparency and the things YouTube should refuse to do.
### Follow-up Questions
- An advertiser asks you to report co-viewers per ad impression. How confident can you be, and how would you express that uncertainty?
- Your model says weekend evening TV viewing is mostly group viewing. How would you check that it has not simply learned "weekend evening TV"?
- How would you detect group viewing on a phone or laptop, where the device gives far fewer clues?
Overview: A technical product question about how YouTube could tell whether a video is being watched by one person or by a group. It tests choosing a use case, obtaining ground-truth labels, combining device, context and content signals into a calibrated estimate, validating it and respecting viewer privacy.