Composer Classification Using a Note Difference Graph

This paper presents a representation for symbolically encoded musical works referred to as a Note Difference Graph. This graph highlights the relative differences between related notes (pitch difference, onset difference, and temporal gap). Our experiments show that when a Graph Neural Network (GNN) is trained to classify classical composers using this note difference graph, it outperforms a network trained with the representation described by Szeto and Wong in which a graph is constructed by identifying related noted. Our approach achieving a 21% increase in classification accuracy on an imbalanced classical music dataset (Szeto and Wong, 2006). The note difference graph employed in this work is derived from the Szeto and Wong representation. Each node in the note difference graph corresponds to an edge in the Szeto and Wong representation (two connected notes in a piece) and contains information relating to the differences between them. Nodes in the note difference graph are joined by an edge if they share any notes in common. The described representation provides improved classification accuracy and reduced bias when using imbalanced datasets. Given the enhanced classification accuracy achieved by the neural network with our representation, we believe that highlighting relationships between notes provides the network with better opportunities to identify salient features.