Text only print
Full screen print
Seminar on Probability and Statistics
Seminar information archive ～06/13｜Next seminar｜Future seminars 06/14～
| Organizer(s) | Nakahiro Yoshida, Hiroki Masuda, Teppei Ogihara, Yuta Koike |
|---|
2026/06/26
13:30-14:30 Room #122 (Graduate School of Math. Sci. Bldg.)
Prof. Hsin-Hsiung 'Bill’ Huang (School of Data, Mathematical, and Statistical Sciences, University of Central Florida)
Scalable Bayesian Conformal Inference for High-Dimensional Spatiotemporal Zero-Inflated Count Data (English)
https://u-tokyo-ac-jp.zoom.us/meeting/register/UmviZSskR766wD4QoUvP2g
Prof. Hsin-Hsiung 'Bill’ Huang (School of Data, Mathematical, and Statistical Sciences, University of Central Florida)
Scalable Bayesian Conformal Inference for High-Dimensional Spatiotemporal Zero-Inflated Count Data (English)
[ Abstract ]
I will present a Bayesian framework for spatiotemporal count data with excess zeros, overdispersion, and ultrahigh-dimensional covariates. The model combines zero-inflated negative binomial regression, TPBN shrinkage priors for sparse fixed effects, graph-Laplacian or SPDE-type spatial random effects, smooth global time effects, and unit-specific Ornstein--Uhlenbeck SDE random effects. Pólya--Gamma augmentation yields a conditionally Gaussian structure, supporting both blocked Gibbs sampling and scalable structured variational inference. I will also discuss split conformal calibration for discrete predictive sets and an auxiliary LAQ/QMLE perspective for the OU component. Simulation studies and a measles surveillance analysis illustrate calibrated prediction and recovery of latent spatiotemporal structure.
[ Reference URL ]I will present a Bayesian framework for spatiotemporal count data with excess zeros, overdispersion, and ultrahigh-dimensional covariates. The model combines zero-inflated negative binomial regression, TPBN shrinkage priors for sparse fixed effects, graph-Laplacian or SPDE-type spatial random effects, smooth global time effects, and unit-specific Ornstein--Uhlenbeck SDE random effects. Pólya--Gamma augmentation yields a conditionally Gaussian structure, supporting both blocked Gibbs sampling and scalable structured variational inference. I will also discuss split conformal calibration for discrete predictive sets and an auxiliary LAQ/QMLE perspective for the OU component. Simulation studies and a measles surveillance analysis illustrate calibrated prediction and recovery of latent spatiotemporal structure.
https://u-tokyo-ac-jp.zoom.us/meeting/register/UmviZSskR766wD4QoUvP2g
