Many Near-Optimal Interpretable Models for High-Stakes Machine Learning Deployment and Discovery

Abstract

Machine learning models have achieved impressive performance across a wide range of predictive tasks. When applied to high-stakes domains such as healthcare, these systems have the potential to improve human lives and generate new scientific insights. However, the traditional deep learning pipeline limits this potential. Practitioners typically train and deploy a single high-performing black-box model, which makes it difficult to understand the model’s reasoning, detect hidden biases, or correct failures once they are discovered. Moreover, because these models are opaque, it is difficult to gain scientific insight by studying them.This dissertation addresses these challenges by leveraging inherently interpretable models and the Rashomon effect. Rather than producing black-box models, I develop \textit{inherently interpretable} models that are equally performant while providing a transparent reasoning process. I leverage the Rashomon effect -- the existence of many models with near-optimal predictive performance -- to correct model failures by simultaneously identifying many alternative models that achieve similar accuracy while avoiding problematic reasoning. I further show how reasoning across many accurate models can enable scientific discovery.First, I introduce AsymMirai, an inherently interpretable deep learning model for predicting a woman’s risk of developing breast cancer within five years from screening mammograms. The model achieves near state-of-the-art predictive performance while relying on a simple and clinically meaningful measure of localized asymmetry between the breasts.Second, I develop a method for efficiently constructing large sets of near-optimal interpretable image classifiers based on prototypical-part networks. By producing many distinct yet accurate models, this approach allows users to interactively edit model behavior and correct confounded reasoning while maintaining predictive performance.Finally, I introduce UNIVERSE, a framework that uses the Rashomon set to study variable importance in the presence of unobserved confounding. By analyzing how variable importance varies across many near-optimal models, UNIVERSE produces finite-sample bounds that characterize which variables could plausibly drive the underlying data-generating process.Together, these contributions demonstrate how interpretability and the Rashomon effect can be used to build machine learning systems that are easier to understand, correct, and learn from in high-stakes domains.