Online Causal Inference under Interventions: A Modular Bayesian Framework

In many science and engineering tasks, the goal is to learn causal models that accurately describe signal behavior. The gold standard for uncovering causal relationships is to apply external shocks, perturbations, or interventions to the system. Once such causal information is obtained, the next step is to incorporate that information into the inference of the causal model. When data arrive as a stream, the need for efficient inference methods becomes apparent. Unlike existing causal discovery methods, which operate in batch mode or require repeated global re-estimation of all causal mechanisms, our approach updates only the components unaffected by interventions. In this work, we propose a conceptually simple yet highly efficient approach for sequentially learning the system's causal structure as interventions are applied. We demonstrate the effectiveness of our method in the context of linear models, which can be extended to more complex modeling frameworks using standard techniques. Our experiments show that the proposed approach is capable of meeting the requirements for scalable, high-speed causal inference.