A0365
Title: Variable discretization-based screening for high dimensional data
Authors: Ryosuke Motegi - Gunma University (Japan) [presenting]
Yoichi Seki - Gunma University (Japan)
Abstract: In statistical modeling with a vast number of predictors, where the number of predictors can be an exponential order of sample size, variable selection plays an essential role in improving prediction performance and interpretability. In such situations, the analyst often takes a two-step process; removing irrelevant predictors with a response variable, then building a model with the remaining predictors using lasso or others. The former step is called variable screening, and various relevance measures have been proposed. An outlier-robust screening method is proposed that can discover variable pairs with nonlinear relationships by looking at the distribution of a response variable conditioned on the discretized predictors. The effectiveness of the proposed method is examined using simulation and real data.
