Searching for Explanations in Rotated Forests: Examining Explainability in Tree-Based Ensemble Models

Abstract

As AI techniques are applied to more decisions that directly and dramatically affect human lives, such as in the medical or military fields, it is increasingly crucial to know why machine learning models make the classifications and recommendations that they do. Explainable AI (XAI) is an emerging field that seeks to modify existing less-interpretable (black box) models to provide more intuitive explanations, or potentially to explore new models that are easily human-interpretable in their decision-making. In this project, we explore the properties of rotated tree-based ensembles---ensembles of inherently explainable decision trees trained on data with simple but diverse transformations applied. Specifically, we investigate whether rotated tree ensembles have the combination of accuracy and interpretability required of candidates for use in meaningfully explainable decision-making/decision-supporting systems.