A Sparse Regression Mixture Model for Clustering Time-Series

Abstract

Abstract. In this study we present a new sparse polynomial regression mixture model for fitting time series. The contribution of this work is the introduction of a smoothing prior over component regression coefficients through a Bayesian framework. This is done by using an appropriate Student-t distribution. The advantages of the sparsity-favouring prior is to make model more robust, less independent on order p of polynomials and improve the clustering procedure. The whole framework is converted into a maximum a posteriori (MAP) approach, where the known EM algorithm can be applied offering update equations for the model parameters in closed forms. The efficiency of the proposed sparse mixture model is experimentally shown by applying it on various real benchmarks and by comparing it with the typical regression mixture and the K-means algorithm. The results are very promising.