Abstract:
|
I will talk about a recent univariate nonparametric regression technique called trend filtering which is a generalization of total variation denoising, a special case of the fused Lasso and closely related to locally adaptive regression splines. Trend filtering presents a natural way of fitting splines where the knots are selected adaptively based on the data points. It is supposed to have attractive spatial adaptivity properties many of which have not yet been established rigorously. I will present some results on the spatial adaptivity of trend filtering. I will also mention connections to shape constrained regression. This is joint work with Adityanand Guntuboyina, Donovan Lieu and Sabyasachi Chatterjee.
|