An integrated method for causality structure and hidden cause discovery

摘要

When we focus on the causality research of real-life applications, we may face many difficulties to discover the causal structure from observational data accurately and detect hidden cause automatically without any prior knowledge. Motivated by these facts, we present a universal definition and an integrated method (ICIC*) for causality discovery with automatic detection of probable hidden causes. Our method induces causal structure by exogenous variables and clique-like structure (IClique). The cues of hidden cause are obtained when redundant edges have been determined. We distinguish two types of causality, be-to-be (11) causality and not be-to-not be (00) causality. To evaluate our method, experiments of causality discovery compared with several methods based on LUCAS (discrete dataset) and Weblog (continuous dataset) and experiment of hidden causes detection based on LUCAS have been done. The results demonstrate higher performance and stronger robustness.