Clustering of Autoregressive Signals Driven by Non-Gaussian Noise in a System Identification Perspective

Clustering is a widely used tool in data analysis, and in many applications, the data to be clustered are time series. In this paper, we propose a clustering method, k-tARs, in the system identification perspective, which is used for autoregressive signals driven by student-t i.i.d. noise. Our algorithm iterates between two steps: the first updates clustering labels, and the second estimates parameters that model the cluster-specific time series. In the step of parameter estimation, due to the non-Gaussian driven noise, the Iterative Reweighted Least Squares method is employed to address the issue of lack of analytical solutions. Numerical experiments are conducted to assess the effectiveness of the k-tARs algorithm over variant experiment setups, and comparative studies against baseline methods further demonstrate its superior performance.