Katherine Tsai

PhD Candidate

University of Illinois at Urbana-Champaign

Katherine Tsai is currently a PhD candidate in Electrical and Computer Engineering department at University of Illinois at Urbana-Champaign working with Prof. Oluwasanmi Koyejo and Prof. Mladen Kolar. She is fortunate to be a recipient of NSF Graduate Research Fellowship. Her research interests lie broadly at the intersection of highdimensional statistics and optimization. In particular, she is interested in developing efficient algorithms to estimate the statistical properties of time-series data with provable guarantees. She has worked on projects in functional data, graphical models, and matrix factorization.  Previously, she completed master’s degree in Electrical and Computer Engineering at UIUC and undergraduate studies at National Taiwan University where she was fortunate to work with Prof. Homer Chen. She was a research assistant at Institute of Information Science, Academia Sinica, where she worked with Prof. Mark Liao and Prof. Li Su. She spent a summer internship at Intel AI Lab in 2020. 

Latent Multimodal Functional Graphical Model Estimation

Multimodal data are collected from same testing subjects through various acquisition techniques. Learning from multimodal data is of great interest and importance to scientists as it provides additional information and helps facilitate new discoveries in the underlying data generation mechanism. On the contrary, graphical models have long roots in studying the conditional independence property, under single data modality. We propose a new integrative framework to estimate single latent graphical model from multimodal functional data. Specifically, we take the generative modeling perspective by modeling the data generation process with the transformation operators mapping functions from the observation space to the latent space. Our method is built on a new partial correlation operator, which we rigorously extend it from the multivariate setting to the functional setting. We then develop an estimator that simultaneously estimates transformation operators and the latent graph via the functional neighborhood regression. Our estimation procedure is provably efficient and scalable that converges to a stationary point with quantifiable statistical error. Our theory proves the latent graph recovery under mild conditions. Both simulation results and empirical results support the benefits of joint estimation.