Contextual Online False Discovery Rate Control
summary
Introduces a formal model of contextual online false discovery rate control, defines contextual generalized alpha-investing procedures and a weighted LORD subclass, proves FDR control when p-values and context are independent under the null, and gives a mixture-model condition ("informativeness") that yields a provable power gap versus context-free methods. Implements online-learned weight functions (neural nets) and reports empirical gains of 20–30% improvement in true discoveries; provides algorithms, theoretical guarantees, and experiments showing the approach works.
AI-written summary · about
chapters (22)
- 0:00Intro
- 1:06Problem of False Discoveries
- 3:15A Tale of Prefixes
- 4:05Offline Multiple Testing
- 8:27Possible Outcomes
- 8:55False Discovery Rate (FDR)
- 10:56Other Side: Statistical Power
- 16:00Prior Work: Online Multiple Testing
- 17:07Using Contextual Information
- 18:08Some Examples
- 20:58Contextual Online Multiple Testing
- 23:46Generalized Alpha Investing Rules in Pictures
- 24:27Generalized Alpha-investing Rules Mathematically
- 26:11How to Incorporate Contextual Information?
- 28:45Proof Idea
- 29:21Increase in Statistical Power in Online Setting
- 30:17Second Piece: Mixture Model
- 31:56Statistical Power Increase with Weighting
- 34:05Modeling the Weight Function
- 34:21Training the Network
- 34:42Experiments on Synthetic Data
- 34:56Concluding Remarks
related talks
34mEqual Opportunity in Online Classification with Partial Feedback
35mOnline Learning with A Lot of Batch Data
33mOnline Multivalid Learning: Means, Moments, and Prediction Intervals
09mIntegrating Differential Privacy and Contextual Integrity
31mOnline Reinforcement Learning With The Help Of Confounded Offline Data
43mContextual Inverse Optimization: Offline and Online Learning
12mDistribution-Free Robustness
28mPrivacy versus Robustness in Federated Learning: Limits and Algorithms