Practical Multicalibration with MCGrad
summary
Claims a practical post‑processing that enforces subgroup calibration by fitting gradient‑boosted additive corrections to model scores using decision‑tree splits, packaged as MCGrad (pip install mcgrad) and measured by the MCE metric and ECCE. Tutorial shows Colab notebooks, LightGBM back end, isotonic regression baseline, and examples where global sigmas drop (e.g. 8.7 → 1.9) but MCE can initially not improve; large internal runbooks report improvements in 80–90% of jobs and deployment at scale across hundreds of production models.
AI-written summary · about
chapters (9)
- 0:00Multicalibration intro
- 3:36Why multicalibration matters
- 16:33Theoretical guarantees
- 26:33Measuring multicalibration
- 39:09Live MCGrad demo
- 54:43Limitations of methods
- 1:01:04MCGrad algorithm design
- 1:13:23Evaluating MCGrad at Meta
- 1:18:19Summary and open-source
related talks
01h 21mMulticalibration: Towards Fair Decision Making
50mA Multi-Group Approach To Algorithmic Fairness
45mMulticalibration, Universal Adaptability and Causality
47mA Multi-Group Perspective on Fairness
36mThe ML Monitoring Flow for Models Deployed to Production
44mHandling predictive uncertainty in Machine Learning
01h 04mPredicting and optimizing the behavior of large ML models
19mHappyMap: A Generalized Multicalibration Method