Pattern recognition in analysing multivariate extremes

Dr Phyllis Wan, Assistant Professor at the Department of Econometrics, Erasmus University Rotterdam

There are effective tools in unsupervised machine learning to uncover structures in multivariate data, such as principal component analysis, clustering analysis and graphical models. In general, these tools are designed to focus on the centre of the data distribution and cannot be directly applied to the tail, that is, the extreme observations.  In this talk, we discuss the adaptation of these techniques to recognise patterns in multivariate extremes.  We present an example of the extreme river flow in the upper Danube basin and invite the audience to discuss its potential use in weather and climate data.