Causal structure learning (CSL), i.e., the derivation of the causal graphical model representing the functional relationships between variables of a system from observational data, is crucial to many domains. However, while current methods of CSL assume the same data type or a specific family of functional relationships, most real-world scenarios incorporate heterogeneous data, e.g., mixed data. Moreover, as the underlying functional causal model and corresponding consistent model-based dependence measures are mostly unknown in advance the selection of inappropriate methods for CSL yields to incorrect results.
As part of our research on Data-Driven Causal Inference we address these challenges through the introduction of information theoretic statistical methods and the development of a modular pipeline for experimental evaluation of CSL from observational data. We demonstrate the applicability of our approaches on both synthetic and real-world settings with regard to the consistency of the estimated underlying causal structures.