Date/Time
Date(s) - 03/03/2023
3:30 pm - 4:30 pm

Speaker:  Ayesha Ali, University of Guelph

Title:  Regularization for High Dimensional Highly Structured Data

Abstract:
When interest lies in understanding how variables in a complex but highly structured system (e.g., genes in a gene network) are associated with an observed trait or outcome, regularization is often used for model selection.  If the predictor set can be represented by a graph, then this structure set can be leveraged during optimization.  Here, we present doubly sparse regression incorporating graphical structure (DSRIG) for variable selection.  DSRIG encourages sparsity, both at the level of the regression coefficients and at the level of individual contributions in a decomposed representation.  We can also prove model consistency with a finite sample error bound.  The model’s high performance and robustness to errors in the input graph structure are demonstrated through simulation, on both random sparse graphs and scale-free graphs, as well as in analysis of real world data.  The error bound is inversely proportional to the sample size and is improved with higher sample sizes.  Time permitting, we will discuss variable selection methods when we have “tiny n, big p”.

Location:  Hamilton Hall, Room 305

Refreshments will be available in the Hamilton Hall Lounge at 3:00 pm