Weak supervision: a new paradigm for unreliable data
summary
Claims a practical pipeline: combine many cheap labeling functions into a generative model (using Snorkel) to produce probabilistic labels, then train a discriminative model on those labels. Implementation uses about 20 labeling functions and reports +2% recall on fraud detection, plus concerns about evaluating generated labels, choosing thresholds, and the critical role of labeling-function density for success.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 1:09Fraud detection types
- 3:37Traditional fraud label challenges
- 5:03Weak supervision need
- 6:47ML hierarchy with weak supervision
- 9:39Labeling functions in weak supervision
- 14:36Generative vs discriminative models
- 16:23Real-world performance gains
- 18:10Evaluating without true labels
- 21:01Labeling function density role
- 23:27Best practices and open questions
- 24:53Resources and hiring
- 25:52Q&A
- 26:17Explaining with labeling functions
- 27:22Balancing functions and sampling
- 28:06Domain knowledge in labeling
- 29:02Graph labeling for fraud detection
- 30:14Model verification challenges
- 31:36Q&A wrap-up
related talks
51mHandling the problem of hand-labeled training data with data programming...
30mEfficient data labelling with weak supervision
35mNo labels? No problem!
20mGreenfield vs. Brownfield Data Labeling
25mData Labeling as Religious Experience
36mProgrammatically creating and managing training data
23mLearning with Limited Labels & Weak Supervision
58mLabeling a Data Set Using Sublinearly Many Queries