Research

Publications

Dynamic Evacuation Recommendation with Inferred Compliance

2026 65th IEEE Conference on Decision and Control (CDC)

Kaiyan Zheng, Hyesun Chung, Ruiwei Jiang, Xi Jessie Yang

Federated Unlearning with Gradient Descent and Conflict Mitigation

AAAI 2025

Zibin Pan, Zhichao Wang, Chi Li, Kaiyan Zheng, Boqi Wang, Xiaoying Tang, Junhua Zhao

Research

Dynamics of Gradient Descent via Function Basis Decomposition

Independent Research Project

University of Michigan | May 2025 - Present

Under supervision of Prof. Salar Fattahi at the Department of Industrial and Operations Engineering.

Characterizing gradient descent dynamics through function basis decomposition across multiple statistical settings. Extending convergence analysis from the population regime to stochastic and empirical risk minimization settings, revealing structural low-dimensional behavior in high-dimensional learning.

Dynamic Evacuation Recommendation with Inferred Trust

Independent Research Project

University of Michigan | January 2025 - Present

Under supervision of Prof. Ruiwei Jiang at the Department of Industrial and Operations Engineering.

Developed a real-time evacuation route recommendation framework that integrates trust inference with dynamic optimization. Formulated and solved a time-evolving optimization model capturing compliance behavior and network congestion. Demonstrated that trust-aware recommendations outperform non-recommendation baselines across key evacuation performance metrics. Submitted to 2026 IEEE Conference of Decision and Control.

Federated Unlearning with Orthogonal Steepest Descent (FedOSD)

Research Project with Zibin Pan et al.

CUHK(SZ) | May 2024 - August 2024

Conducted research at the Laboratory of Energy Internet under Prof. Junhua Zhao, focusing on privacy-preserving optimization in federated learning systems.

Co-designed and implemented the FedOSD federated unlearning algorithm, introducing a stabilized unlearning loss and SVD-based orthogonal steepest descent direction to mitigate gradient conflicts. Verified minimal computational overhead of orthogonalization and conducted comprehensive empirical evaluations under heterogeneous client distributions. Contributed to the AAAI 2025 manuscript.