Kernel-Based Causal Discovery in Large Time Series Datasets

Discovering causal relationships from a large-scale observational time series dataset is an emerging topic in the current field of research, such as neuroscience and system monitoring. Traditional causal inference approaches in such fields are challenging since the curse of dimensionality and the large sample size. Here, we propose a kernel-based causal discovery (KCD) method by capitalizing on a characterization of conditional independence in reproducing kernel Hilbert spaces. Specifically, KCD uses an iterative and heuristic procedure to add variables into the causal parents set and then adopts the kernel-based conditional independence test to determine the true causal parents. We validated this approach on a large-scale synthetic dataset that simulates the typical properties of real-world data. The results of the experiments have shown that the proposed method is superior to currently existing technologies in terms of detection capabilities, which provides a completely new way to explore how to discover and quantify causal networks from time series in the current research field.a