Modern enterprise systems - and IT systems generally - generate enormous amounts of data related to the execution of workflows and other processes. We can harness that data using process mining techniques, and related machine learning algorithms, to provide insight into real-world process execution, and develop predictive models for process enhancement, risk mitigation and anomaly detection. This thesis explores approaches to learning two contrasting classes of models, namely imperative and declarative models, and hybrids of the two. Roughly speaking, the former captures processes as explicit flows or procedures, whereas the latter captures the most essential constraints, rules or properties of a process and is often encoded in a logic resembling natural language. The aim is to harness the strengths of both to learn accurate and interpretable models. With a strong focus on one-to-one empirical comparisons across the modeling paradigms, the project aims to ground hybrid process mining on a solid statistical basis and frame it as a classical inference task.