Abstract:
|
We have extended the binary predictive modeling technique, Feature Augmentation via Nonparametrics and Selection (FANS), to the ordinal response setting by employing proportional odds boosting. Proportional odds boosting additively combines many weak (but potentially flexible) models with small degrees of freedom to form a single strong learner. We specified these weak learners as linear models of augmented features, where an augmented feature is defined using the ordinal class-conditional kernel density estimates of the original predictors. Furthermore, we incorporated the FANS approach of data splitting and prediction averaging in order to make efficient use of the data. Several proportional odds boosting models are fit using different versions of the original data, and the models are aggregated by combining the predicted probabilities. We present the results of a simulation study as well as an analysis of a microarray gene expression dataset to compare our method to several competing methods.
|