Survival forest with partial least squares for high dimensional censored data

Random forest and partial least squares have proved wide applicability in numerous contexts. However, the combination of these versatile tools has seldom been studied. Inspired by a relatively new decision tree ensemble called rotation forest, we introduce a new survival ensemble algorithm using partial least squares regression and the Buckley-James estimator within the framework of random forest. First, the approach taken to cope with the high dimensionality is to reduce the dimension by a random subspace method. Then, censored survival times are imputed by the Buckley-James estimator. After dimension reduction and time imputation, partial least squares regression is applied to extract the features. Similar to rotation forest, all extracted components are used as covariates in a bagged survival tree to predict the survival probabilities. Experimental results on a variety of simulation and real datasets demonstrate that the proposed approach is a strong competitor to other popular survival prediction models under high or ultra-high dimensional setting.