Modeling and Predicting Song Adjacencies in Commercial Albums

This paper examines whether latent structure may be discovered from commercially sold albums using features characterizing their song’s adjacencies. We build a largescale dataset from the first 5 songs of 8,505 commercial albums. The dataset spans multiple artists, genres, and decades. We generate a training set (Train) consisting of 11,340 True song adjacencies and use it to train a mixture of multivariate gaussians. We also generate two disjoint test sets (Test1 and Test2), each having 11,340 True song adjacencies and 45,360 Artificial song adjacencies. We perform feature subset selection and evaluate on Test1. We test our model on Test2 in a standard retrieval setting. The model achieves a precision of 22.58%, above baseline precision of 20%. We compare this performance against a model trained and tested on a smaller dataset and against a model that uses full-song features. In the former case, precision is better than the large scale experiment (24.80%). In the latter case, the model achieves precision no better than baseline (20.13%). Noting the difficulty of the retrieval task, we speculate that using features which characterize song adjacency may improve Automatic Playlist Generation (APG) systems.