A0369
Title: On the ability of random forests to model interactions
Authors: Constanze Lehner - University of Passau (Germany) [presenting]
Abstract: It is often argued that random forests can implicitly account for interactions due to the tree-like structure of its base learner. The binary recursive partitioning approach of classification and regression trees divides the observations into more homogeneous subgroups by making sequential univariate splits on covariates. Recent contributions suggest that such univariate partitioning schemes may not be appropriate to adequately capture the effect of interrelated covariates on the response variable. To explore these arguments about whether interactions are modeled in random forests, we use simulated datasets with different interaction patterns and evaluate the predictive performance and variable importance measures of random forests. Our analysis also provides a review of interactions in the random forest literature, from their origins in the decision tree literature to current applications. For the review, we discuss interaction concepts in different research areas in advance.
