Theodore Kypraios

will speak on

Bayesian Detection and Prediction of Disruptive Events using Twitter

Time: 3:00PM
Date: Thu 25th April 2024
Location: E0.32 (beside Pi restaurant) [map]

Abstract: The volume of tweets on Twitter is increasing exponentially, thus
providing us with numerous opportunities for detecting the occurrence of
major events in real-time. We develop a state-space model for detecting
disruption on the National Railway in Great Britain in a timely fashion,
by using the content and volume of tweets referring to delays and
disturbance in the railway. A time-inhomogeneous Poisson process,
$\lambda(t)$, is proposed to model the number of tweets whose
time-dependent intensity function is parameterized such that it captures
the observed periodic pattern in the data. A hidden Markov process that
represents the state of the railway through time ('normal'/ 'abnormal')
then modulates the Poisson process. We further extend the model allowing
for dependence in the data by developing multivariate Markov-modulated
Poisson Process and discuss how to overcome the inferential challenges
posed by such a model.

We develop a computationally efficient MCMC algorithm to learn the
parameters governing $\lambda(t)$ and infer the state of the railway
through time by utilizing a Forward-Backward algorithm to efficiently
update the unobserved process. We demonstrate through extensive
simulation studies that (i) we can successfully recover the model's
unknown parameters, (ii) predict the unobserved state with high accuracy
with robust results towards model misspecification. Finally, we
illustrate via Bayesian filtering how to predict the future state of
railway in real-time given the observed number of tweets.

(This talk is part of the Statistics and Actuarial Science series.)

PDF notice

Return to all seminars