Please Note: This seminar will be given online.
University of North Carolina at Greensboro
Link to join seminar: Hosted on Zoom
Conquering the potentially crippling effect of too many interactions on the Gauss-Dantzig selector
Screening designs are used when, with limited resources, important factors are to be identified from a large pool of factors. Typically, a screening experiment will be followed by a second experiment to study the effect of the identified factors in more detail. As a result, the screening experiment should ideally screen out a large number of factors to make the follow-up experiment manageable, without screening out important factors.
The Gauss-Dantzig Selector (GDS) is often the preferred analysis method for screening designs. While there is ample empirical evidence that fitting a main-effects model can lead to incorrect conclusions about the factors if there are interactions, including two-factor interactions in the model increases the number of model terms dramatically and challenges the GDS analysis. I will discuss a new analysis method, called Gauss-Dantzig Selector – Aggregation over Random Models (GDS-ARM), which aggregates the effects from different iterations of the GDS analysis performed using different sets of randomly selected interactions columns each time. The GDS-ARM draws its motivation in part from random forests by building many models, and by identifying important factors after running a GDS analysis on all of these models. I will discuss the proposed method, study its performance, and elaborate on potential future directions.