Multilevel Spectral Clustering for extreme event characterization
Abstract
Direct spectral clustering framework was first proposed to extract general pattern events within multivariate time series. This study investigated the way to identify extreme events, i.e. short duration and/or particular events, with no assumption about their emission date, duration and/or shape. A Multilevel Spectral Clustering (M-SC) architecture is proposed and compared with state-of-the-art clustering methods from a simulated manually labeled time series. Due to these promising empirical results, this new deep architecture is applied on marine field data.