survivalRKHS
Model-free Variable Selection in Reproducing Kernel Hilbert Space for right-censored data
M. Matabuena and P. Félix and C. Meixide-Garcı́a
Abstract. A novel variable selection method is presented for those sort of regression problems in survival analysis where the response variable can be right-censored. The proposed approach is model-free, that is, no model assumption is made between the response and pre- dictors, and is implemented through a two-stage procedure: firstly, a previous independence screening is performed to efficiently reduce dimensionality in high-dimensional data sets; and secondly, from the remaining variables, the informative predictors are identified as those showing a gradient function substantially different from zero. Both stages are formulated in a learning framework based on Reproducing Kernel Hilbert Space (RKHS). The effectiveness of this method is supported experimentally by multiple synthetic data sets and a real data set from survival in cancer patients, proving that the new method outperforms classical methods in scenarios with non-linear relation- ships between variables.