Title: Soft clustering of time series: New methods considering fuzzy, mixture models and probabilistic-D approaches
Authors: Jose Vilar - Universidade da Coruna (Spain) [presenting]
Borja Lafuente-Rego - Universidade da Coruna (Spain)
Abstract: New soft clustering algorithms to deal with time series considering fuzzy, nonparametric mixed models and probabilistic D-clustering techniques are introduced. In all cases, dissimilarity between series is measured in terms of squared Euclidean distance between sample quantile autocovariance vectors (QAD). The fuzzy model relies on a fuzzy C-medoids approach where QAD is used to calculate distances between series and medoids. Based on the asymptotic representation of the log-periodogram by means of a nonparametric regression model with log-exponentially distributed errors, an EM algorithm is used to estimate the components of a finite mixture model involving nonparametric approximations of the log-periodograms for each cluster. The probabilistic D-clustering technique is directly extended to the time series framework by considering QAD as a suitable feature-based distance between series. Supported by the nice properties of QAD, all the proposed algorithms show a very good clustering performance in simulations. Furthermore, the procedures are compared and the strengths and weaknesses of each one remarked and discussed.