How is drift detection an unsupervised problem?
mediumAnswer
- Compare current feature distributions to training / baseline distributions without labels.
- Methods: (1) KS test per feature (univariate).
- (2) Population Stability Index (PSI).
- (3) Multivariate: MMD (Maximum Mean Discrepancy), classifier drift test (train binary classifier reference vs current — AUC > 0.7 = drift).
- (4) Wasserstein / Jensen-Shannon divergence.
- Standard in production ML monitoring — Datadog, Arize, WhyLabs, Fiddler all implement variants.
Check yourself — multiple choice
- Only supervised
- Compare current vs baseline distributions unsupervised: KS / PSI / MMD / classifier drift test / Wasserstein — foundation of ML monitoring platforms
- Random
- Same as accuracy
Drift detection: KS / PSI / MMD / classifier drift — unsupervised distribution comparison.
#anomaly-detection#applications
Practise Unsupervised Learning
214 interview questions in this topic.
Related questions
- Anomaly detection in time series — what changes?
- Covariate drift vs concept drift vs label drift — the differences.
- Scan statistics — when do you use them?
- CUSUM change-point detection — how does it work?
- How do you handle 99.9% normal / 0.1% anomaly training data?
- How do you explain why a point is anomalous?