How would you predict a car’s turning intention?
Company: Meta
Role: Data Scientist
Category: Machine Learning
Difficulty: medium
Interview Round: Onsite
At an intersection, there are **n vehicles** stopped or approaching.
For each vehicle, you have a short history (e.g., last 3–10 seconds at 10 Hz) of:
- **Position** (x, y) in a map-aligned coordinate system
- **Velocity** (vx, vy) and optionally acceleration
- (Optional) lane/map context (lane centerlines, stop lines)
You want to predict each vehicle’s **intention** over the next few seconds as one of:
- **Left turn**
- **Go straight**
- **Right turn**
**Task:** Describe how you would build, train, and evaluate an ML model for intention prediction.
In your answer, cover:
1. Labeling strategy and time horizon (when is intention defined?).
2. Feature engineering vs sequence models; how to incorporate map/lane geometry.
3. How to handle multi-agent interactions (other cars influence intent).
4. Evaluation metrics (accuracy vs calibration vs early prediction; per-class metrics; cost-sensitive errors).
5. Handling class imbalance, distribution shift across intersections/cities, and intent changes at the last second.
6. How you would estimate uncertainty and use it in downstream planning.
Quick Answer: This question evaluates applied machine learning skills for vehicle intention prediction, covering temporal sequence modeling versus feature engineering, map- and lane-aware representations, multi-agent interaction reasoning, labeling/time-horizon definition, evaluation metrics and calibration, class imbalance and distribution shift handling, and uncertainty estimation for downstream planning. It is commonly asked because intention prediction is safety-critical in autonomous driving and probes a candidate’s ability to make system-level design trade-offs across data labeling, model architecture, training/validation, and risk-aware evaluation. Category/Domain: Machine Learning; Abstraction level: system-level applied ML for data scientist roles (mid-to-senior).