Time Series Causal Discovery Using a Hybrid Method

In this paper, we introduce a novel framework, Neural-HATS for inferring causal structures in time series data using a hybrid method. Neural-HATS uniquely combines conditional independence (CI) testing with continuous optimization-based learning methods to enhance causal discovery. Specifically, it leverages an attention-based encoder-decoder architecture with Kernel Conditional Independence (KCI) testing to enable direct CI tests between time series. These CI test results are integrated into continuous optimization algorithms, enhancing both causal inference accuracy and the effectiveness of continuous optimization models. Experimental evaluations demonstrate that Neural-HATS achieves improved causal graph accuracy.