Can we create a better model for time series data? This paper introduces a new statistical model that segments time series data into regimes with approximately linear dynamics and learns the parameters of each regime. This model combines hidden Markov models and linear dynamical systems. The authors present a variational approximation that maximizes a lower bound on the log-likelihood, utilizing both forward and backward recursions. Tests on artificial and natural data suggest the viability of variational approximations for inference and learning in switching state-space models.
Published in Neural Computation, this paper aligns with the journal's emphasis on computational methods, machine learning, and neural networks. The research presents a novel statistical model and a variational approximation for time series analysis, contributing to the field of machine learning. Its relevance to neural computation and data analysis makes it suitable for the journal's readership.