Comparative Study of Survival Support Vector Machine and Random Survival Forest in Survival Data

Abstract

Survival analysis is a statistical procedure in analyzing data with the response variable is time until an event occurs (time-to-event). In the last few years, many classification approaches have been developed in machine learning, but only a few considered the presence of time-to-event variable. Random Survival Forest and Survival Support Vector Machine are machine learning approach which is a nonparametric classification method when dealing with large data and a response variable of survival time. Random Survival Forest is tree based method that using boostrapping algorithm, and Survival Support Vector Machine using hybrid approaches between regression and ranking constrain. The data used in this study is generated data in the form of right-censored survival data. This study uses the RandomForestSRC and SurvivalSVM packages on R software. This study aimed to compare the performance of the Survival Support Vector Machine and Random Survival Forest methods using simulation studies. Simulation results on right-censored survival data using binary predictor variables scenario indicate that the Survival Support Vector Machine (SSVM) method with Radial Basic Function Kernel (RBF Kernel) has the best model performance on data with small volumes, whereas when the data volume becomes larger, the method that has the best performance is Survival Support Vector Machine using Additive Kernel. Meanwhile, Random Survival Forest is a method that has the best performance for all conditions in mixed predictor variables scenario. Method, proportion of censored data and size of data are factors that affect the model performance.