Method

Weak supervision

Programmatic or heuristic labels combined into training data.

9 cases from 9 companies apply Weak supervision. 2 were checked in an online A/B test; 2 report results in production.

Typical tasks: Catch fraud and abuse, Detect anomalies, Rank search results, Moderate content at scale and 4 more. Companies: Stripe, Zillow, Samokat.tech, Celsus, DoorDash, Netflix and 3 more. Often combined with: Threshold selection, Multi-label classification, Autoencoder anomaly detection, Isolation Forest, Gradient boosting and 35 more.