An Adaptive Weighted Deep Survival Forest

Abstract

An adaptive weighted deep survival forest model for survival analysis is proposed. It can be regarded as an extension of the adaptive weighted deep forest. First, it is based on the deep forest being an ensemble-based model including a set of the random forests organized in a special form of levels of a forest cascade similarly to layers in neural networks. Second, it is based on introducing a special scheme of assigning the weights to instances in the deep forest, which allows us to adapt the random survival forests at every level to training data. One of the main ideas underlying the proposed model is to introduce and to apply a marginal concordance index (MC-index) as a measure of an instance performance and to compute the weights as functions of the MC-index. Numerical examples with real data illustrate the proposed adaptive model.