Abstract
Efficient methodologies capable of capturing influential effects count as a key concept of medical prognosis when predicting the time until the occurrence of an event of interest. As a major concern, high-dimensional feature spaces need to be addressed in survival analysis. To this end, we propose a novel ensemble framework based on semi-stochastic selection of features that can be used for survival prediction in the presence of high-dimensional feature spaces. A sequential ensemble learning scheme is presented in which each survival tree is trained on a subspace of features. We incorporate a Bayesian framework with a conjugate prior for selecting features in each step by a semi-stochastic procedure. To explore the most influential features in the learning process, the Beta-Bernoulli bandit scheme is utilized. For interpretation, posterior means provided by the Bayesian framework are used as a measure of variable importance. We evaluate the performance of our proposed method based on real data analysis, and the efficiency of the novel variable importance measure by simulation studies.