Title:Fitting High-Dimensional Interaction Models with Error Control

Abstract:There has recently been a renewed interest in polynomial regression despite its many well-known problems: higher-order polynomials are often unstable near the boundaries of the feature space, interaction spaces are often highly correlated, and even in rather small data sets the number of interactions is unwieldy. Standard feature selection issues remain such as how to determine a final model which generalizes well. This paper solves these problems with a sequential algorithm for searching the feature-space of higher-order interactions which is based on the principle of marginality. Furthermore, the procedure controls a notion of false rejections by implementing Revisiting Alpha-Investing. This provides a valid stopping criterion to prevent over-selection.