Procedural content generation (PCG) is a growing area of research focused on leveraging artificial intelligence in the design and creation of content (e.g., levels, environments, stories, etc.) oftentimes for video games. However, most current PCG approaches are domain specific or require a substantial amount of domain knowledge to be used across multiple domains. We want to determine whether more general approaches to PCG are possible (i.e., approaches that can be applied across large classes of domains without customization or domain knowledge). The first key contribution of this dissertation is to show that machine learning approaches, specifically Markov models, can be used to model and generate levels across multiple domains by replacing domain knowledge with training data, while still being able to capture much of the domain information, such as structural level information and player interactions. The second key contribution of our work is a new theoretical framework to understand PCG approaches based on machine learning, and provide a unifying view of this new class of approaches, highlighting similarities, differences, and providing insights into future avenues of research. Our third main contribution is the development of extensions to these machine learning-based approaches that allow for more control over the generated content and more accurate modeling of the given domain.