Interpretable Classification via Supervised Variational Autoencoders and Differentiable Decision Trees

Eleanor Quint, Garrett Wirka, Jacob Williams, Stephen Scott, N.V. Vinodchandran

Abstract: As deep learning-based classifiers are increasingly adopted in real-world applications, the importance of understanding how a particular label is chosen grows. Single decision trees are an example of a simple, interpretable classifier, but are unsuitable for use with complex, high-dimensional data. On the other hand, the variational autoencoder (VAE) is designed to learn a factored, low-dimensional representation of data, but typically encodes high-likelihood data in an intrinsically non-separable way. We introduce the differentiable decision tree (DDT) as a modular component of deep networks and a simple, differentiable loss function that allows for end-to-end optimization of a deep network to compress high-dimensional data for classification by a single decision tree. We also explore the power of labeled data in a supervised VAE (SVAE) with a Gaussian mixture prior, which leverages label information to produce a high-quality generative model with improved bounds on log-likelihood. We combine the SVAE with the DDT to get our classifier+VAE (C+VAE), which is competitive in both classification error and log-likelihood, despite optimizing both simultaneously and using a very simple encoder/decoder architecture.

TL;DR: We combine differentiable decision trees with supervised variational autoencoders to enhance interpretability of classification.

Keywords: interpretable classification, decision trees, deep learning, variational autoencoder