Visualizing States by Autoregressive Gaussian Process Dynamical Models

摘要

Gaussian process dynamical models (GPDMs) are a nonlinear dimensionality reduction technique for time series that provides a probabilistic representation of time series in terms of Gaussian process priors. In this paper, we study an extension of GPDM to visualize states of time series. Conventional GPDM associates a state with an observation value, and therefore cannot represent observations changing over time by one state. As a result, since the state changes every time the observation value changes, the GPDM approach creates a difficult to understand visualization of state transition. To overcome the problem, we propose an autoregressive GPDM called ARGPDM, which associates a state with a vector autoregressive (VAR) model and therefore can represent observations changing over time by one state. The ARGPDM approach creates an easier to understand visualization since the state changes only when the VAR model changes. We demonstrate experimentally that the ARGPDM approach provides better visualization than conventional GPDM.