B0929
Title: Building random forest explanations through a locally accurate rule extractor
Authors: Celine Vens - KU Leuven (Belgium) [presenting]
Abstract: Random forests are machine learning methods characterised by high performance and robustness to overfitting. However, since multiple learners are combined, they are not as interpretable as a single decision tree. A novel method is proposed, which is building explanations through a locally accurate rule extractor (Bellatrex), which is able to explain the forest prediction for a given test instance with only a few diverse rules. Starting from the decision trees generated by a random forest, the method selects a subset of the rules used to make the prediction, represents them as a vector, clusters the vectors, and picks a rule from each cluster to explain the instance prediction. The effectiveness of Bellatrex is tested on 89 real-world datasets and the validity of the method is demonstrated for binary classification, regression, multi-label classification and time-to-event tasks. It is deemed that it is the first time that an interpretability toolbox can handle all these tasks within the same framework. It is also shown that Bellatrex is able to approximate the performance of the corresponding ensemble model in all considered tasks, and it does so while selecting at most three rules from the whole forest. Finally, a comparison with similar methods in the literature also shows that the proposed approach substantially outperforms other explainable toolboxes in terms of predictive performance.
