Causal inference for relationships between time series data

Abstract

Learning of causal relations in time series data is important to investigate the causes of abnormalities and predict sensor values that cannot be measured directly. Traditional methods for the problems widely use Vector Autoregressive (VAR) models in many fields but can be applied to stationary time series data only. This study proposes a method to reveal correlations and time-lags in non-stationary time series data by combining Gaussian Process Dynamic Models (GPDM) and Multi-Task Gaussian Process (MTGP). The results of applying the method to climate open data are also shown.