Research & Projects
🔬 Research Projects
Robustness of Multiple Knockoff Aggregation Methods
- Advisor: Prof. Lijun Wang, School of Mathematical Sciences, Zhejiang University
- Period: Sept. 2025 – Present
- Status: First-author manuscript in preparation
- Keywords: False Discovery Rate (FDR) · Multiple Knockoffs · Heavy-tailed Distribution · Feature Collinearity · Copula
Innovation & Contributions:
- Systematically evaluated the robustness of state-of-the-art aggregation methods (e-value based, based-voting, Multiple Data Splitting) under extreme heavy-tailed distributions, identifying severe power breakdowns (Power \(\to 0\)) under Cauchy noise.
- Diagnosed failure mechanisms within the single thresholding step, demonstrating that overly strict thresholds lead to power loss, and empirically investigated strategies (e.g., relaxing FDR control thresholds) to rescue selection power.
- Analyzed the impact of feature collinearity and different feature importance statistics (LASSO path difference, Random Forest), revealing that derandomized knockoffs recover their superior FDR-control performance at lower correlations (\(\rho = 0.1\)).
- Investigated the parameter sensitivity of the \(\alpha_{kn}/\alpha_{ebh}\) ratio, exploring empirical CDF approximation frameworks to dynamically calibrate selection thresholds.
- Researched Copula-based extensions (e.g., Latent Gaussian and Vine Copula Knockoffs) to address and relax the strict distributional assumptions of standard Model-X procedures.
Uncertainty Quantification via Conformal Prediction
- Advisor: Prof. Guanhua Chen, School of Medicine and Public Health, UW-Madison
- Period: Mar. 2026 – Present
- Status: In progress
- Keywords: Conformal Prediction · Distribution-Free Inference · Uncertainty Quantification
Innovation & Contributions:
- Investigating conformal inference frameworks to construct distribution-free predictive intervals for complex statistical models.
💻 Selected Computational Projects
C-Based CNN Forward Inference Engine
- Period: Spring 2025
- Tech Stack:
C,Model Optimization,Binary I/O
Key Implementations:
- Implemented a convolutional neural network (CNN) forward inference engine completely from scratch in C.
- Achieved a 75% latency reduction via rigorous channel pruning while maintaining \(\sim 99\%\) classification accuracy.
Uncertainty Quantification for PDEs in Random Domains
- Period: Fall 2025
- Tech Stack:
Python,Stochastic Galerkin,Karhunen-Loève Expansions
Key Implementations:
- Reproduced numerical solutions for elliptic Stochastic Partial Differential Equations (SPDEs) utilizing stochastic mappings.
- Applied Legendre polynomial chaos and Karhunen-Loève expansions for robust uncertainty quantification in complex domains.