Title: Ordinal forests: Prediction and covariate importance ranking with ordinal response variables
Authors: Roman Hornung - Institute for Medical Information Processing, Biometry and Epidemiology, University of Munich (Germany) [presenting]
Abstract: The ordinal forest method is a random forest-type prediction method for ordinal response variables. The trees in ordinal forests are regression trees that use optimized score values in place of the ordered class values of the response variable. The optimization of the score values aims at maximizing the estimated prediction performance of the forest. Ordinal forests allow prediction using both low-dimensional and high-dimensional covariate data and can additionally be used to rank covariates with respect to their importance for prediction. An extensive comparison study using several real datasets and simulated data reveals that ordinal forests tend to outperform competitors in terms of prediction performance. Moreover, it is seen that the covariate importance measure currently used by ordinal forest discriminates influential covariates from noise covariates at least similarly well as the measures used by competitors. The rationale underlying ordinal forests of using optimized score values in place of the class values of the ordinal response variable is in principle applicable to any regression method beyond random forests for a continuous outcome that is considered in the ordinal forest method. While the original ordinal forest algorithm only also to perform class point predictions, a recent update allows predicting class probabilities.