I am Chudi Zhong. I am a Ph.D. student in computer science at Duke University, advised by Prof. Cynthia Rudin. I am also working closely with Prof. Margo Seltzer on my research projects. I recieved my B.S. in Statistics and Information Science from UNC-Chapel Hill and M.S. in Statistical Science from Duke.

My primary research interest lies in the intersection of machine learning, optimization, and human-model interaction. My goal is to develop machine learning systems that are easier to troubleshoot, interact with, and gain knowledge from. Specifically, I work on optimizing interpretable models and exploring their Rashomon sets.

Sparse Decision Tree Optimization

We developed an optimal sparse decision tree algorithm called Generalized and Scalable Optimal Sparse Decision Trees (GOSDT). It combines the branch-and-bound and dynamic programming methodology with custom bit-vector libraries and computation reuse. This algorithm is different from greedy splitting and pruning, such as used in CART and C4.5, and it is able to produce trees that can surprisingly often be comparable with black-box models in terms of prediction accuracy.

To further improve the scalability of the optimal classification tree algorithms, we proposed a set of techniques that decreases the time of producing optimal sparse trees by two to three orders of magnitude. This allows us to produce high-performing interpretable models on much larger and more challenging datasets than ever before.

Sparse Generalized Additive Models

We presented a fast and scalable algorithm to train generalized additive models by directly solving an ℓ₀ regularized logistic regression problem. We proposed a set of techniques for a coordinate-descent-and-local-swap-based algorithm. These techniques allow our algorithm can handle thousands of observations and features in minutes and can achieve accuracy and AUC comparable to black-box models on real datasets.

Rashomon Set of Interpretable Models

A set of well-performing models is known as the Rashomon set. This term came from the Rashomon Effect coined by Leo Breiman, which describes the fact that many equally good models explain the same data well. In this project, we aim to propose a new framework for trustworthy machine learning through constructing a set of near-optimal interpretable models and providing users with unprecendented flexibility to visualize, explore, select, and modify not only one but many well-performing models.

We have developed algorithms to efficiently and accurately construct the Rashomon set of sparse decision trees (TreeFARMS) and sparse generalized additive models and tools to facilitate human-model interactions using the Rashomon set.