Causal Discovery from high-dimensional observational data

With the rise of digital observational data, there has been an increasing amount of attention to the discovery of causal relations from large datasets. In the last three decades two major approaches have emerged to deal with high-dimensional Causal Discovery; well-known SL methods optimized by bypassing their exponentially growing conditioning tests, and quasi-experimental designs equipped with machine learning algorithms to efficiently search for promising hypotheses. These methods have mainly focused on dealing with the computational complexity and the assumptions made about the data. In this thesis, the goal is to expand the use of these approaches, by attempting to solve various types of problems that are inspired by real-world applications, and are hard or impossible to solve by the general causal discovery methods. First is the Relational Causal Discovery, in which the prior knowledge of the association between variables can be used to improve accuracy and reliability. Second is the Stratified Causal Discovery, which identifies causes for different subpopulations, and potentially different underlying mechanisms. Third is the Causal Profile Discovery, which specifies the temporal sequence of causes to have the most significant effect. Fourth is the Compound Causal Discovery, which identifies the sets of causes that are only jointly sufficient. The thesis will be concluded by discussing the applications of the proposed methods and how they can be used as platforms for other potential problems.