Abstract:In many real applications, such as climate, finance and social media among others, we are often interested in extreme events. An important part of modeling extremes is discovery of covariates on which the quantities related to the extremes are dependent, as this may lead to improved understanding and the discovery of new causal drivers of extremes. Despite developments in sparse covariate discovery algorithms, adaptations to extremes can fail because the tail attributes do not follow a Gaussian distribution. In this paper, we proposed a sparse Bayesian framework for discovery of covariates that influences the frequency of extremes based on the Poisson description of extremes frequency and hierarchical Bayesian description of a sparse regression model. We developed an efficient approximation algorithm based on the variational Bayes approach to estimate the distribution over regression coefficients that indicate dependence of extremes on the corresponding covariates. Experiments with synthetic data demonstrate the ability of the approach to accurately characterize dependence structures. Applications to rainfall extremes suggest new insights relevant for improved understanding of hydrological extremes under climate variability and change.
Enter your feedback below and we'll get back to you as soon as possible.