Tabu-Based Causal Discovery for Time Series with Variable Lags

A newly developed algorithm for deriving causal Bayesian networks from time-series data tackles the issue of varying time lags between causes and their effects. Current techniques usually operate under the assumption of a constant lag window, which limits their ability to identify dependencies that manifest at different intervals. The innovative Tabu-based structure learning algorithm seeks a time-sequenced directed graph, ensuring that each edge adheres to temporal order while accommodating edge-specific lags up to a defined maximum. By employing a decomposable BIC-based scoring system with node-specific effective sample sizes, this method enhances traditional approaches by optimizing lag values for each edge, thus facilitating more precise causal discovery in intricate dynamic systems. This research is available on arXiv and marks a notable advancement in causal inference from temporal data.