Séminaire
Organisme intervenant (ou équipe pour les séminaires internes)
INRAE
Nom intervenant
Guillem Rigaill
Résumé
In recent years, many methods have been proposed for detecting one or multiple changepoints offline or online in data streams. The reason for such a keen interest in changepoint detection methods lies in its importance for various real-world applications, including bioinformatics, climate and oceanography, econometrics, or finance. The increasing volume of data streams poses significant computational challenges for detecting changepoints online. Likelihood-ratio based methods are effective, but their straightforward implementation becomes impractical online.
In the first part of this presentation, I will detail how one would maximize likelihood ratio statistics in a straightforward manner in quadratic time, which is too slow in an online setting. I will then present some informal arguments and simple simulations to conjecture that these calculations could be done faster without sacrificing exactness (that is, without resorting to computational heuristics).
In the second part of the presentation, I will explain how one can indeed maximize likelihood ratio statistics faster (provably quasi-linear in expectation) and exactly by leveraging a link between changepoint detection and geometry. This link holds true for a surprisingly large range of univariate and multivariate models: essentially, it suffices that the likelihood ratio can be algebraically written as one from the natural exponential family. I will conclude the presentation with a few simulations and, if time permits, an application to an NBA dataset.
This is joint work with Liudmila Pishchagina, Gaetano Romano, Paul Fearnhead and Vincent Runge, and based on the following arxiv: https://arxiv.org/abs/2311.01174
In the first part of this presentation, I will detail how one would maximize likelihood ratio statistics in a straightforward manner in quadratic time, which is too slow in an online setting. I will then present some informal arguments and simple simulations to conjecture that these calculations could be done faster without sacrificing exactness (that is, without resorting to computational heuristics).
In the second part of the presentation, I will explain how one can indeed maximize likelihood ratio statistics faster (provably quasi-linear in expectation) and exactly by leveraging a link between changepoint detection and geometry. This link holds true for a surprisingly large range of univariate and multivariate models: essentially, it suffices that the likelihood ratio can be algebraically written as one from the natural exponential family. I will conclude the presentation with a few simulations and, if time permits, an application to an NBA dataset.
This is joint work with Liudmila Pishchagina, Gaetano Romano, Paul Fearnhead and Vincent Runge, and based on the following arxiv: https://arxiv.org/abs/2311.01174
Lieu
Amphi C2.0.37
Date du jour
Date de fin du Workshop