UCLA

We study the task of explaining machine learning classifiers. We explore a symbolic approach to this task, by first compiling the decision function of a classifier into a tractable decision diagram, and then explaining its behavior using exact reasoning techniques on the tractable form. On the compilation front, we propose new algorithms for encoding the decision functions of Bayesian Network Classifiers and Binarized Neural Network Classifiers into tractable decision diagrams. On the explanation front, we examine techniques for generating a variety of instance-based and classifier-based explanations on tractable decision diagrams. Finally, we evaluate our approach on real-world and synthetic classifiers. Using our algorithms, we can efficiently produce exact explanations that deepen our understanding of these classifiers.