Designing a Machine Learning System to Flag Inauthentic User Reviews
Company: Microsoft
Role: Product Manager
Category: Machine Learning
Difficulty: medium
Interview Round: Other
How would you use machine learning to detect fake reviews on a platform where users post ratings and written reviews?
```hint Look past the text
One review on its own often reads like any other. Consider what the reviewer's account history, the timing of reviews on a listing, and links between accounts can reveal, and where you would get trustworthy labels to learn from.
```
### Clarifying Questions
- Which surface are the reviews on: a product marketplace, an app store, or local business listings, and can a review be tied to a verified purchase or download?
- Which kinds of fake are in scope: paid or incentivized praise, reviews by the seller or a competitor, bot-generated text, or coordinated rings of accounts?
- What can the system do with a flagged review: block it at submission, hold it for moderators, remove it, or keep it visible but leave it out of the average rating?
- Which error costs more here: a fake review that stays up, or a genuine customer's review taken down?
### What a Strong Answer Covers
- Why fake reviews hurt shoppers, honest sellers and the platform, and a precise definition of the abuse types the system targets
- Where labels come from when ground truth is scarce and biased, and which signals beyond review text (account, behavior, timing, account networks) carry the most weight
- A layered detection design: submission-time rules, a per-review model and detection of coordinated groups, with score thresholds mapped to actions and human review
- How success is measured offline and in production, including precision at each action threshold, the prevalence of missed fakes, and moderator workload
- How the system copes with adversaries who adapt, plus appeals for reviewers and sellers
### Follow-up Questions
- Fraudsters switch to language models that write varied, natural-sounding reviews. Which parts of your system still work, and what would you add?
- A seller says dozens of genuine five-star reviews disappeared after a successful promotion. How do you investigate, and what would you change?
- How do you show leadership the system is working when you never directly observe the fake reviews you missed?
Overview: A product-focused machine learning question about catching fake reviews on a platform where users post ratings and written reviews. It tests how a candidate defines review fraud, finds labels when ground truth is scarce, combines text, account and network signals, ties model scores to actions, and measures both precision and missed abuse.