An $L_0$-Norm Regularized Method for Multivariate Time Series Segmentation

Abstract

A multivariate time series segmentation model based on the minimization of the negative log-likelihood function of the series is proposed. The model is regularized by the $L_0$-norm of the time series mean change and solved by an alternating process. We use a dynamic programming algorithm in order to determine the breakpoints and the cross-validation method to find the parameters of the model. Experiments show the efficiency of the method for segmenting both synthetic and real multivariate time series.