Adaptive Path of Prediction: An Unsupervised Method for Modeling Note-Level Informational Hierarchy of Polyphony

Polyphonic music presents a unique challenge for computational modeling due to the complex interactions of multiple simultaneous musical streams and the need to capture both local and global structural relationships. We propose Adaptive Path of Prediction, a discrete diffusion model that learns the informational hierarchy of polyphony in an unsupervised manner. By training the model to find optimal note-removal paths, and to reversibly reconstruct these selectively removed notes, we reveal how critical musical events—that sustain to later stages of data corruption—maximize the preserved information and guide the prediction of remaining content. Drawing on compression learning theory, we posit that such adaptively-discovered "anchor notes" reflect the system's ability to make an explicit abstraction of polyphonic music. Our experiments demonstrate that the model converges on consistent note-importance distinctions and can achieve better reconstruction performance in selected denoising paths than random ones. Furthermore, the model's assignment of note importance during the training process increasingly aligns with a reductive music analysis dataset, suggesting that our unsupervised framework can uncover structural hierarchies consistent with established music-theoretical views.