Variable importance for scattered data,
Prof Pierre Gremaud
The determination of the relative importance of the inputs to a black box function Y is a fundamental task: not only does it yield an improved understanding of the underlying problem, it also facilitates the development of efficient predictors. For many applications, evaluating Y is computationally expensive; for some applications, Y may only be accessed through a given dataset. Global sensitivity analysis is thus often performed on surrogate models, introducing bias. We propose a new notion of variable importance and demonstrate its performance when used in conjunction with surrogate models from non-parametric statistics.
These concepts and methods will be introduced from scratch and illustrated by multiple examples.
Joint work with Joey Hart.
This seminar will be hosted in Erskine 244.
Prof Gremaud will also run a workshop on Tuesday 24 November: Uncertainty Quantification: introduction and examples.
Prof Pierre Gremaud comes to us from North Carolina State University, his research interests include numerical PDEs, mathematical modeling, material and medical science applications, granular materials.
You are warmly invited to attend this seminar. Please email email@example.com to register your interest.