I will overview recent algorithms developed in the Prediction Analysis Lab at MIT for interpretable predictive modeling. I will discuss applications to some of society's most critical problems in healthcare and energy grid reliability. In particular:
1) Interpretable Linear Models, where the coefficients of the models lie in a discrete set for interpretability. This is being applied to predict sleep apnea.
2) The Latent State Hazard Model, which learns a decomposition of the hazard rate into two terms. One of these terms represents mechanical degradation and it is monotonically increasing, whereas the other term does not have this restriction. This is applied to wind turbine maintenance.
3) The Bayesian Case Model, which is a prototype model for clustering. Each cluster is centered around one of the observations (a prototype), and both the prototypes and the important subspaces for each cluster are learned from data.
|
ASA Meetings Department
732 North Washington Street, Alexandria, VA 22314
(703) 684-1221 • meetings@amstat.org
Copyright © American Statistical Association.