Skip to main content
stat
Statistical and Algorithmic Guarantees for High Dimensional Single Index Model (HDSIM)
Add to Calendar 2021-04-02T14:10:00 2021-04-02T15:00:00 UTC Statistical and Algorithmic Guarantees for High Dimensional Single Index Model (HDSIM)
Start DateFri, Apr 02, 2021
10:10 AM
to
End DateFri, Apr 02, 2021
11:00 AM
Presented By
Ran Dai (University of Nebraska Medical Center)
Event Series: SMAC Talks

Single index model (SIM) is an important semi-parametric extension of linear models. In this ongoing work, we consider a shape constrained high dimensional SIM (hdSIM), where the link function is monotonic (but not necessarily smooth) and the coefficient parameter is high dimensional and sparse. We propose a scalable projection-based iterative approach, namely  the  “orthogonal  gradient single-index model” (OG-SIM) algorithm, which alternates between sparse-thresholded orthogonalized gradient steps and isotonic regression steps to recover the coefficient vector. We show theoretical guarantees for the estimation of both the link function and the coefficient parameter. Our theoretical work will be based on very mild assumptions on the design matrix X and error term. In particular, X can be fixed or random. If X is random, it can be asymmetric and the error distribution can depend on X. We study the performance of our algorithm via both numerical studies and also an application on a rocker protein sequence data. This is joint work with Hyebin Song, Rina Foygel Barber, and Garvesh Raskutti.