Vine copula-based Bayesian classification for multivariate time series of electroencephalography eye states

Sometimes classification tasks have to be based on multivariate time series data collected for each class. In these situations the data for each class might exhibit non-stationary behaviour together with complex dependence structures. We propose a vine copula-based approach to capture these features in each class before applying a Bayesian classifier. Vine copulas have been very successful in modelling asymmetric tail dependence among variables and are coupled with non-stationary univariate time series to model the multivariate time series data for each class. We illustrate this classification approach using data from a neural activity experiment using electroencephalography, where we want to classify the eye state. The level of neural activity was collected over time for multiple locations on the scalp. Our approach is able to identify relevant locations and allows for a model-based interpretation of the data generating process. A cross-validation study with comparison to competitor classifiers for this data set shows good performance of the proposed classifier.