Evaluate Guangzhou performance with limited data
Company: WeRide
Role: Data Scientist
Category: Machine Learning
Difficulty: medium
Interview Round: Technical Screen
##### Question
You have built an autonomous-driving evaluation system using a large amount of labeled data from Beijing. Now the company wants to operate in Guangzhou. You do not want to rebuild the entire evaluation pipeline from scratch, and you can only collect a small amount of Guangzhou data.
How would you evaluate whether the autonomous-driving system is likely to perform well in Guangzhou under this limited-data setting? Your answer should address:
1. How to define the target population and success criteria for Guangzhou.
2. How to assess whether the Beijing-based evaluation system (metrics, tooling, calibration, thresholds) transfers to Guangzhou.
3. What kinds of distribution shift to look for, and how to detect shift between Beijing and Guangzhou.
4. How to use the small Guangzhou sample efficiently, and how to combine the large Beijing dataset with the small Guangzhou dataset.
5. How to quantify uncertainty and make a go / no-go recommendation when the evidence is limited.
**Follow-up:** Suppose Guangzhou contains important scenarios that are rare or absent in Beijing—for example different road topology, scooter/two-wheeler density, weather, traffic-agent mix, signage, local driving behavior, or map quality. How should that change your data-collection strategy? Be explicit about what to sample and how to prioritize edge cases.
Quick Answer: Evaluate autonomous-driving performance transfer from Beijing to Guangzhou with limited local data, covering target population, distribution shift, scenario taxonomy, stratified sampling, importance weighting, partial pooling, uncertainty, rare scenarios, and staged launch decisions.