Computational Mathematics and Statistics Seminar by Yifei Xiong: Inference from Privatized Data: Data Augmentation MCMC and Flow-Based Posterior Approximation

Speaker: Yifei Xiong,Ph.D. candidate, Purdue University

 

Title: Inference from Privatized Data: Data Augmentation MCMC and Flow-Based Posterior Approximation

 

Abstract: Differential privacy enables data sharing while protecting sensitive information, but it also makes inference substantially more difficult. This talk presents two complementary approaches to inference from privatized data. The first is a data augmentation MCMC approach, in which the unobserved confidential data are introduced as latent variables and updated jointly with the model parameters, with recent improvements aimed at making the imputation step more efficient. The second is a simulation-based approach that uses normalizing flows to learn posterior or likelihood approximations from simulated privatized data in settings where the likelihood is intractable. Together, these methods provide two distinct computational routes to inference under privacy constraints.

If you’re unable to attend the seminar in person, we also have a Teams link available.

 

 

 

Computational Mathematics and Statistics Seminar

Link