Canada Spends Logo

Explorez notre analyse du budget 2026

Grant: Discovery Grants Program - Individual

University of Waterloo

$27,000.00

Multivariate Probability Predictions and Time-Varying Networks

EF 2023-24

Résumé
In recent years, I have been studying different types of interactions, such as genetic epistasis and those between protein side chains. This proposal will consider not just static interactions but also those that vary over time. It is perhaps not entirely unfair to say that we statisticians have largely been left out of the recent "deep learning" movement. But I have found some interesting ways for us to take advantage of it as well. Together with a colleague and a PhD student, we can train neural networks to generate samples from fairly arbitrary probability distributions. This has allowed us to produce some interesting statistical applications---such as generating quasi Monte Carlo samples approximately from "almost any" distribution, and making probability predictions for multivariate time series without restricting ourselves to only a few parametric copula families to describe the underlying cross-sectional dependence. When making these multivariate probability predictions, we took advantage of the famous Sklar decomposition of a multivariate distribution into its marginals and a copula. Each marginal was allowed to depend on the past, while the copula---which captures the interactions between different dimensions---was assumed to be stationary over time. A neural network was then used to learn the copula in a fully flexible manner. We were conscious of the fact that holding the copula stationary was restrictive for many applications, and removing such a restriction would be desirable, but doing so would make the neural network much harder to train. I will be tackling this difficulty next. After we make any predictions, whether point or probability predictions, it is natural for us to want to evaluate them, e.g., on held-out test data. It turns out that assessing multivariate probability predictions is still very much an open problem. There are conflicting messages about different scoring rules, and the community has not yet reached a consensus. I would like to crack some of these puzzles. I will also continue my work on dynamic network data. Together with a PhD student and a co-supervisor, we have extended the popular stochastic block model (SBM) to deal with network interactions that come on and off in continuous time, and further adapted the resulting "continuous-time SBM" to analyze basketball games. But there are still many unsolved problems. For example, we may observe that network connections are changing over time, but different factors may be driving these changes---the effects of the covariates, connection probabilities between different communities, and node-memberships in different communities could all be varying over time. But when multiple factors are involved, there are intricate identifiability problems, and sensible constraints will be needed. By joining my research program, doctoral, master's, and undergraduate students will be able to develop various quantitative skills that are much needed in today's economy.
Programme
Discovery Grants Program - Individual
Emplacement
Ontario, Canada
Type de bénéficiaire
Academia
Type
Grant
Source
NSERC's Awards Data

Tel que publié

Application id
RGPIN-2023-03337
Name nom
Zhu, Mu
Department département
Statistics and Actuarial Science
Organization id
33
Institution établissement
University of Waterloo
Province en
Ontario
Province fr
Ontario
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
27000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1508
Committee name en
Mathematics and Statistics
Committee name fr
Mathématiques et statistique
Area of application code
1200
Area of application group en
Advancement of knowledge
Area of application group fr
Avancement des connaissances
Area of application en
Advancement of knowledge
Area of application fr
Avancement des connaissances
Research subject code
9999
Research subject group en
Not available
Research subject group fr
Not available
Research subject en
Non disponible
Research subject fr
Non disponible
Application title
Multivariate Probability Predictions and Time-Varying Networks
Keyword
copula; dependence; energy score; neural network; probability prediction; stochastic block model; variogram score
Application summary
In recent years, I have been studying different types of interactions, such as genetic epistasis and those between protein side chains. This proposal will consider not just static interactions but also those that vary over time. It is perhaps not entirely unfair to say that we statisticians have largely been left out of the recent "deep learning" movement. But I have found some interesting ways for us to take advantage of it as well. Together with a colleague and a PhD student, we can train neural networks to generate samples from fairly arbitrary probability distributions. This has allowed us to produce some interesting statistical applications---such as generating quasi Monte Carlo samples approximately from "almost any" distribution, and making probability predictions for multivariate time series without restricting ourselves to only a few parametric copula families to describe the underlying cross-sectional dependence. When making these multivariate probability predictions, we took advantage of the famous Sklar decomposition of a multivariate distribution into its marginals and a copula. Each marginal was allowed to depend on the past, while the copula---which captures the interactions between different dimensions---was assumed to be stationary over time. A neural network was then used to learn the copula in a fully flexible manner. We were conscious of the fact that holding the copula stationary was restrictive for many applications, and removing such a restriction would be desirable, but doing so would make the neural network much harder to train. I will be tackling this difficulty next. After we make any predictions, whether point or probability predictions, it is natural for us to want to evaluate them, e.g., on held-out test data. It turns out that assessing multivariate probability predictions is still very much an open problem. There are conflicting messages about different scoring rules, and the community has not yet reached a consensus. I would like to crack some of these puzzles. I will also continue my work on dynamic network data. Together with a PhD student and a co-supervisor, we have extended the popular stochastic block model (SBM) to deal with network interactions that come on and off in continuous time, and further adapted the resulting "continuous-time SBM" to analyze basketball games. But there are still many unsolved problems. For example, we may observe that network connections are changing over time, but different factors may be driving these changes---the effects of the covariates, connection probabilities between different communities, and node-memberships in different communities could all be varying over time. But when multiple factors are involved, there are intricate identifiability problems, and sensible constraints will be needed. By joining my research program, doctoral, master's, and undergraduate students will be able to develop various quantitative skills that are much needed in today's economy.