USING PAIRWISE LINK PREDICTION AND GRAPH ATTENTION NETWORKS FOR MUSIC STRUCTURE ANALYSIS

The task of music structure analysis has been mostly addressed as a sequential problem, by relying on the internal homogeneity of musical sections or their repetitions. In this work, we instead regard it as a pairwise link prediction task. If for any pair of time instants in a track, one can successfully predict whether they belong to the same structural entity or not, then the underlying structure can be easily recovered. Building upon this assumption, we propose a method that first learns to classify pairwise links between time frames as belonging to the same section (or segment) or not. The resulting link features, along with node-specific information, are combined through a graph attention network. The latter is regularized with a graph partitioning training objective and outputs boundary locations between musical segments and section labels. The overall system is lightweight and performs competitively with previous methods. The evaluation is done on two standard datasets for music structure analysis and an ablation study is conducted in order to gain insight on the role played by its different components.