Learning dynamic causal mechanisms from non-stationary data

Causal discovery from non-stationary time series is an important but challenging task. Most existing non-stationary approaches only consider the changes of causal coefficients, which are merely satisfied in real-world scenarios. In this paper, we introduce a Gaussian-based Variational Temporal Abstraction model (GVTA) to detect and learn non-stationary causal mechanisms from multiple time series. First, we utilize a hierarchical cyclic state-space model to detect the stationary states from the non-stationary time series. Second, we use the Gaussian process algorithm to estimate the causal mechanism for each stationary state. Experimental results on both simulation and real-world data demonstrate the correctness and effectiveness of our method.

Automated summary

In this paper, we propose a Gaussian-based Variational Temporal Abstraction model (GVTA) to detect and learn non-stationary causal mechanisms from multiple time series. Our method utilizes a hierarchical cyclic state-space model to detect stationary states and uses a Gaussian process algorithm to estimate the causal mechanisms for each stationary state. Experimental results demonstrate the correctness and effectiveness of our method.