Explaining a Random Survival Forest by Extracting Prototype Rules

Tree-ensemble algorithms and specifically Random Survival Forests (RSF) have emerged as prominently powerful methods for survival data analysis. Tree-ensembles are very accurate, robust, resilient to overfitting, and can naturally handle missing values as well as categorical data. However, since they consist of multiple models, they are not as interpretable as single decision trees. In this work, we propose a method that learns to extract a limited number of representative rulesets from the ensemble providing explanations of the ensemble model's outcome. We propose a local approach, focusing on explaining predictions for a specific sample, and is mainly divided into three parts; tree-filtering, low dimensional representation, and prototype ruleset extraction. Here, we employ RSF as the ensemble model but our approach is generalised to other settings as well. We conducted preliminary experiments on both binary classification using relevant data as well as time-to-event predictions in a survival analysis context. The obtained results demonstrate that our approach performs comparably well to the original Random (Survival) Forest that it explains, while based only on few trees from the whole forest.