B1021
Title: Fitting prediction rule ensembles with multiply-imputed data, and adaptive and relaxed lasso penalties.
Authors: Marjolein Fokkema - Leiden University (Netherlands) [presenting]
Abstract: Prediction rule ensembling (PRE) aims to derive interpretable regression and classification models, with accuracy similar to that of tree ensembles but better interpretability. The RuleFit algorithm is a flexible method for PRE, which originally proposed the use of lasso regression for obtaining a sparse final rule ensemble. There is a lack of evidence on how to best deal with multiply-imputed data with PRE. While pooling provides the most promising avenue in terms of predictive accuracy, it is likely detrimental to interpretability. The performance of stacking and pooling approaches is compared in terms of accuracy and interpretability. Furthermore, since the introduction of RuleFit, the relaxed and adaptive lasso penalties have been proposed which promise to offer better stability, sparsity and/or accuracy than the original lasso. Their performance is assessed in terms of accuracy and sparsity in complete and multiply-imputed data.
