GRACE Enhances Causal Discovery in High-Dimensional Time Series

Identifying causal relationships within high-dimensional time-series data, such as climate patterns or gene regulation, poses a significant challenge for researchers. Traditional constraint-based methods, while statistically rigorous, struggle with the computational demands of nonlinear conditional independence (CI) tests at scale. Score-based alternatives, on the other hand, often require arbitrary thresholds to convert continuous edge scores into binary decisions, leading to less robust results. A new approach called GRACE (Gated Refinement for Accurate Causal Edge discovery) addresses these limitations by integrating constraint-based discovery with a novel gated refinement mechanism. GRACE employs Hard Concrete gates with L0 regularization, which produce distinct, bimodal distributions for candidate edges, making the binary decision of causality much more robust than methods relying on L1 or attention. It starts with a fast linear CI skeleton to generate high-recall candidates, then uses a single gated model to prune false positives by learning which edges genuinely improve prediction. Extensive experiments on synthetic datasets with various topologies and dimensions up to 100 variables demonstrated GRACE's superior performance, significantly improving F1 scores and precision while maintaining high recall. It also proved substantially faster than expensive nonlinear CI tests. In a real-world application to river flow data, GRACE successfully recovered a high percentage of known causal edges with minimal false positives, even under challenging conditions like rainfall confounders and distributional shifts.