[ABE-L] Insper Data Science Academic Seminar - Helio Migon

[Hedibert Lopes]

Data Science

Academic Seminar


March 24, 2022

12:00 p.m. de São Paulo, Brasil (UTC/GMT -03:00)

https://zoom.us/j/97410173464



Speaker: Helio Migon

University: Universidade Federal do Rio de Janeiro - UFRJ


Title: k-PARAMETRIC DYNAMIC GENERALIZED LINEAR MODELS


 Abstract:

 Dynamic generalised linear models may be seen simultaneously as an
extension to dynamic linear models and to generalised linear models,
formally treating serial auto-correlation inherent to responses observed
through time.  The current presentation revisits the Dynamic Linear Model
class, focusing on specific concerns such as monitoring and intervention
analysis. An approach based on information geometry is used to extend the
DGLM to the k-parametric exponential family. The suggested technique
supports multinomial responses and may be extended to accommodate
compositional responses on k = d + 1 categories, while maintaining the
sequential feature of the Bayesian inferential procedure, resulting in
real-time inference. The updating scheme  takes advantage of the
exponential family's conjugate structure, which ensures computing
efficiency. The method is based on concepts like Kullback-Leibler
divergence and the projection theorem, putting it in line with modern
approaches to variational inference. We will show applications to
real-world data to demonstrate the applicability of the proposed strategy.
The method's computational efficiency is demonstrated in comparison to
other approaches, as well as its flexibility to quickly accommodate new
information when strategically needed, all while maintaining aspects of
monitoring and intervention analysis, as well as discount factors, which
are common in sequential analyses.
-------------- Pr�xima Parte ----------
Um anexo em HTML foi limpo...
URL: <http://lists.ime.usp.br/pipermail/abe/attachments/20220323/e81968a6/attachment.htm>