The Influence of Loss Function on Oblique Survival Tree Induction

Survival trees are a common machine learning tool designed to handle censored data, where only partial information about failure events is available. Most survival tree models work by recursively dividing the feature space using splits defined by single attributes in internal nodes. However, there is a less common type known as oblique survival trees, which use more attributes to create splits in the form of hyperplanes. In this paper, we depart from the typical top-down approach and focus on globally induced oblique survival trees, aiming to optimize both prediction accuracy and model complexity. We propose using two different loss functions— the integrated Brier score and a likelihood-based loss— in the process of oblique survival tree induction. We then compare the resulting models in terms of their predictive performance and complexity.