Title: Model-based co-clustering of multivariate time-dependent data
Authors: Alessandro Casa - Free University of Bozen-Bolzano (Italy) [presenting]
Charles Bouveyron - INRIA, Universite Cote d'Azur (France)
Elena Erosheva - University of Washington (United States)
Giovanna Menardi - University of Padova (Italy)
Abstract: Multivariate time-dependent data arise when multiple features are measured for a set of units over different time instants. When dealing with such data, flexible statistical tools are needed in order to account for characteristics such as the relations among both time observations and variables, the possible subject heterogeneity and arbitrarily shaped time evolutions. A new co-clustering strategy is outlined, grouping simultaneously variables and individuals, and being adequate both for longitudinal and functional data. The proposed approach relies on the shape invariant model which is embedded in the latent block model, representing the most popular model-based co-clustering strategy. To account for the specific features of the shape invariant model, the estimation procedure is carried out by means of a suitable modification of the SEM-Gibbs algorithm. The resulting methodology flexibly introduces different, and possibly user-defined, notions of cluster and, by partitioning matrices into homogeneous blocks, provides parsimonious representations of high-dimensional and complex structured time-dependent data. Lastly, the explicit modelling of time evolutions allows for meaningful interpretations of the clusters.