Introduction of AUC-based splitting criteria to random survival forests
Fabian Eifler · Open access LMU (Ludwid Maxmilian's Universitat Munchen) · 2014
The goal of this thesis is to implement a new, AUC based splitting criterion for random survival forests which was inspired by the Concordance Index and to evaluate the performance of that criterion versus the already established log rank splitting.In the first part, the methodological background along with the theory behind the new splitting rule is introduced.Alongside that the implementation of all functions used in this thesis is described in its parameters, return values as well as functionality.The second part of the thesis consists of evaluating the new criterion on two real life datasets as well as two artificially generated ones.The performance is evaluated on different aspects and tuning parameters such as forest size, number of covariables selected in each split, size of the terminal nodes, censoring rate and sensitivity to noise.The results of the evaluation show that the new splitting criterion is on average performing slightly better than the already established one at the cost of a higher computational effort.However with high censoring rates or lots of noise in the dataset the established log rank splitting criterion starts to have problems with delivering meaningful results whereas the new splitting criterion still works well.