Machine Learning and Music Theory: Models for Hierarchical Music Generation and Analysis

Abstract

Recent advancements in deep learning have significantly improved automatic music generation, yet many models remain uninterpretable and struggle to capture fundamental musical features such as hierarchical rhythmic and harmonic-melodic structure. A lack of transparency in these models limits their practical usability and prevents meaningful human interaction with the generated music. Interpretability is crucial not only for understanding and refining generative processes but also for ensuring that models adhere to established musical principles.To address these challenges, this dissertation introduces novel generative frameworks that prioritize interpretability while maintaining musical coherence. We propose a probabilistic approach that models music composition using human-informed hierarchical structures based on Schenkerian analysis (SchA), ensuring that generated melodies and harmonizations align with particular compositional styles. Additionally, we introduce a graph-based representation for SchA, providing a structured and scalable method for encoding hierarchical musical relationships. Finally, we develop deep learning models that formulate hierarchical music analysis as graph pooling and link prediction problems, leveraging graph neural networks.Our results demonstrate that incorporating domain knowledge and hierarchical representations leads to more transparent and controllable generative models. These methods produce music that is structurally coherent, aesthetically compelling, and more interpretable than existing deep learning approaches. By bridging the gap between computational models and human music analysis, this research paves the way for future advancements in AI-assisted composition, interactive music generation, and computational musicology.