• causal abstraction
• causality
• machine learning

By providing a formal framework for decision-making and what-if reasoning, the concept of causation could fundamentally shape how artificially intelligent agents interact and reason with an environment. In the last decades, the graphical approach to causality, where variables are represented by nodes and their edges stand for their causal relations, has attracted significant research and gained a large popularity. Despite a plethora of methods dedicated to the problem of recovering these causal graphs from data, their application to datasets composed by a large number of variables is still a pressing issue. Causal Abstraction is a recently defined framework that enables concise representations of large systems through significantly smaller graphical causal models. These abstract models retain causal properties of the system by aggregating the higher-dimensional representation. Overall, the thesis tackles different open issues in the context of learning causal abstractions from data. By doing so, we report original contributions for structure learning, i.e., the problem of recovering a graphical structure from data, theory of causal abstraction, and finally causal abstraction learning for linear causal models.