Optimal survival trees ensemble

Selection of accurate and diverse trees based on individual and collective performance in an ensemble has recently been studied for classification and regression problems. Following this notion, the possibility of selecting optimal survival trees is considered in this work. Initially, a large set of survival trees are grown by the method of random survival forest. Using out-of-bag observations for each corresponding survival tree, the trees grown are ranked in ascending order with respect to their prediction errors. A certain number of the top ranked survival trees are selected to be assessed for their collective performance in an ensemble. An ensemble is initiated from the top ranked selected survival tree and further trees are tested one by one by adding them to the ensemble. A survival tree is selected for the final ensemble if it improves the performance by assessing on an independent training data. This ensemble is called optimal survival trees ensemble (OSTE). The proposed method is checked on 17 benchmark datasets and the results are compared with those of random survival forest, conditional inference forest, bagging and Cox proportional hazard model. In addition to improved predictive performance, the proposed method also reduces the number of survival trees in the ensemble as compared to the other tree based methods. Furthermore, the method is implemented in an $R$ package called "OSTE''.