Effect of Interaction Terms on Recursive Partitioning Techniques (Recursive Partitioning, Unbiased Conditional Inference Trees, Random Forest): Some Experiments with the Icreater Functions

Abstract

The idea of generating and searching for interactions and transformations is implemented in R. The icreater function adds log, sqrt, and negative reciprocal transformations suggested by Tukey and then the data set with these transformation is used to generate NXN interactions and this final data set is fed into recursive partitioning, conditional inference trees and random forests. This approach is tested for 3 credit data sets: German, Brazillian credit card data set and home equity data set. Adding transformed interaction terms improves predictive accuracy in 2 out of the 3 data sets. So it is shown to work sometimes.