Local causal structure learning for streaming features

Abstract

Local causal structure learning (LCSL) is the process of discovering and distinguishing direct causes (parents) and direct effects (children) of a given target variable (T) without learning a global causal structure. However, state-of-the-art LCSL algorithms can only address static feature space, while ignoring the dynamic changes of feature space with streaming features. For high dimensional data, existing methods fail to efficiently distinguish causality between features and the target variable. To address these issues, we propose Local Causal Structure Learning for Streaming Features (LCSLSF). LCSLSF comprises two sequential steps, as it first dynamically mines causal features to construct an approximate Markov Blanket (aMB) of the target variable. It proceeds by filtering irrelevant and redundant features, retaining causal features as much as possible from streaming features. Second, it learns local causal structures by mining and splicing the V-structure of feature nodes in batches from aMB of the target variable. On benchmark Bayesian networks with the number of variables ranging from 20 to 724, LCSLSF achieves a significantly better performance than its rivals. We demonstrate its effectiveness by conducting a real-world case study on causal discovery in medical diagnostics. Code is released at https://github.com/youdianlong/LCSLSF.