Dynamics of Musical Success : A Bayesian Nonparametric Approach

We model the dynamics of musical success of albums over the last half century with a view towards constructing musically well-balanced playlists. We develop a novel nonparametric Bayesian modeling framework that combines data of different modalities (e.g. metadata, acoustic and textual data) to infer the correlates of album success. We then show how artists, music platforms, and label houses can use the model estimates to compile new albums and playlists. Our empirical investigation uses a unique dataset which we collected using different online sources. The modeling framework integrates different types of nonparametrics. One component uses a supervised hierarchical Dirichlet process to summarize the perceptual information in crowd-sourced textual tags and another time-varying component uses dynamic penalized splines to capture how different acoustical features of music have shaped album success over the years. Our results illuminate the broad patterns in the rise and decline of different musical genres, and in the emergence of new forms of music. They also characterize how various subjective and objective acoustic measures have waxed and waned in importance over the years. We uncover a number of themes that categorize albums in terms of sub-genres, consumption contexts, emotions, nostalgia and other aspects of the musical experience. We show how the parameters of our model can be used to construct music compilations and playlists that are likely to appeal to listeners with different preferences and requirements.