Spatial stochastic process clustering using a local a posteriori probability

This paper addresses the problem of spatial stochastic process clustering in a model-based framework. A data set is used, in which each realization has two components : a non-uniform time series, which describes the process evolution with independent increments and a set of additional attributes which describes the system characteristics. It is assumed that realizations with similar additional attributes tend to have the same cluster label. The aim is to find out the unknown cluster labels and the parameters of the statistical model characterizing the processes. Thus this is a kind of a problem of spatial clustering with a time component. The proposed method is based on an EM procedure and takes into account the proximity of the additional attributes using a local a posteriori probability. The importance of the neighborhood influence is tuned thanks to a parameter. The method is illustrated using simulated data.