Zakk Heile

Hi — I’m Zakk Heile, a Computer Science & Mathematics undergraduate at Duke University, supported by the Benjamin N. Duke Merit Scholarship.

My current research focuses on optimization problems for interpretable machine learning. In many high-stakes settings, optimal sparse models can be competitive with black-box models while remaining interpretable, which is critical for responsible decision making and useful for machine learning practitioners as a way to understand model behavior and motivate principled feature engineering. I am particularly interested in approximating Rashomon sets, which are the sets of all optimal and near-optimal models within a given model class. Rashomon sets are useful for characterizing variable importance across all near-optimal models, for making predictions without imputing missing values, and for analyzing disagreements between equally good models, providing a more complete picture of uncertainty and structure than any single model. My latest work develops metaheuristic algorithms to approximate Rashomon sets by performing rollout on subproblems, where candidate splits are evaluated using single decision tree algorithms that themselves select splits via rollout of cheaper decision tree algorithms.

Previously, I have worked on geometric data analysis problems, including rigid-invariant distances for comparing distributions, point clouds, and shapes; on interpretable natural language processing methods for healthcare applications in traumatic brain injury; and on real-time particle collision filtering for deployment on custom FPGA-based hardware at CERN’s Large Hadron Collider.