[Ml-stat-talks] Fwd: Wilks Statistics Seminar: Alessandro Rinaldo, Today, April 1st @ 12:30pm, Sherrerd Hall 101
bee at princeton.edu
Fri Apr 1 10:38:35 EDT 2016
Talk of interest today.
---------- Forwarded message ----------
*** Wilks Statistics Seminar ***
DATE: Today, April 1, 2016
LOCATION: Sherrerd Hall 101
SPEAKER: Alessandro Rinaldo, Carnegie Mellon University
TITLE: A Framework for Distribution Free Regression
ABSTRACT: We develop a general framework for distribution free regression
using the idea of conformal inference. The method converts any pre-chosen
estimator of the regression function to produce prediction bands with
finite sample validity. The resulting pre- diction band maintains the good
theoretical properties of the original estimator under standard
assumptions, while the finite sample validity is guaranteed under
essentially no assumptions. We develop several computational and inference
techniques under this framework that are broadly applicable to both modern
high-dimensional setting and classical settings. First, we develop
easy-to-compute in-sample prediction bands with both finite-sample and
asymptotic coverage guarantee. Second, we use our framework to define a new
notion of variable importance that does not require any model assumptions.
This leads to a new, predictive-based method for making inferences about
the variable inference called LOCO (Leave-One-Covariate-Out). Third, we
show how to extend our approach to produce prediction bands with varying
local width, which is suitable for non-constant noise variance. We also
apply the method to nonparametric smoothing, tree regression and sparse
additive models. This is joint work with Max G’Sell, Jing Lei, Ryan
Tibshirani, and Larry Wasserman.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Ml-stat-talks