About me
I am a Ph.D. student in Statistics & Data Science at Carnegie Mellon University, advised by Prof. Shixiang Zhu. My research focuses on uncertainty quantification, particularly conformal prediction, and human–AI collaboration for decision-making.
Before joining Carnegie Mellon, I received an M.S. in Statistics from the University of Southern California, where I worked with Prof. Paromita Dubey on change-point detection for metric-space data. I also earned an M.S. in Statistics from the University of Wisconsin–Madison, where I worked with Prof. Kris Sankaran and Prof. Keith Levin on machine-learning interpretability and statistical network analysis. I received B.S. degrees in Statistics and Applied Mathematics from Renmin University of China.
Research interests
- Conformal prediction and distribution-free uncertainty quantification
- Reliable and secure machine learning
- Scientific machine learning and statistical decision-making
- Human–AI collaboration
Recent work
Beyond Predicting Responses: Conformal Inference for Latent Distributional Parameters
We introduce LatentCP, a prior-free framework that constructs uncertainty sets for unobserved, instance-specific distributional parameters using only observed context–response pairs and a specified forward model. It provides finite-sample marginal coverage without latent calibration labels or knowledge of the latent mixing distribution, with an application to latent wildfire intensity.
Learning to Test: Physics-Informed Representation for Dynamical Instability Detection
We develop a test-oriented, physics-informed representation for monitoring dynamical systems under changing operating conditions. The method converts deployment-time safety monitoring into a distributional hypothesis test with controlled Type I error, avoiding repeated large-scale simulation.
Generative Conformal Prediction with Vectorized Non-Conformity Scores
We propose a generative conformal framework that uses vectorized non-conformity scores to form adaptive prediction regions for complex, multidimensional outcomes. Density-ranked uncertainty balls allocate coverage more efficiently while retaining statistical validity.
Change Point Inference for Non-Euclidean Data Sequences using Distance Profiles
We develop a tuning-parameter-free scan statistic for detecting and locating distributional changes in metric-space data. The method has asymptotic guarantees and applies broadly to objects such as distributions and networks when pairwise distances are available.
See the publications page for a complete list and Google Scholar for citation records. For research correspondence, email minxingz@andrew.cmu.edu.