DP Mixture of Warped Correlated GPs for Individualized Time Series Prediction

Kian Ming A. Chai, Yun Jie Serene Yeo, Weiping Priscilla Fan, Si Hui Maureen Lee, Junxian Ong, Poh Ling Tan, Yu Li Lydia Law, Kok-Yong Seng

Abstract: We propose a Dirichlet process (DP) mixture of warped Gaussian processes (GP) to model a collection of profiles, each of which is a set of correlated time-series. Given a new profile, we predict one of its time-series given the others in a streaming manner. Variational inference is used for tractability. This includes an online formulation of the projected process approximation of GPs to handle real-time prediction. We demonstrate our method on predicting the body-core temperature of subjects given their heart-rate and skin temperature.