Graph embedding of music structures for machine learning approaches

Abstract

Several works on representation learning for graph-structured data have been proposed in recent literature. However, most of such techniques have several downsides. On the one hand, graph kernels which use handcrafted features (e.g., shortest paths) are hampered by poor generalization problems. On the other hand, methods for learning representations of whole graphs deal with unattributed or single-attributed graphs.In this work, we propose a novel technique for graph embedding learning able to take into account multi-attribute graphs (from 1 to an arbitrary number). Given a multi-attribute graph, the proposed method generates an embedding vector as follows: (i) the graph is split into several single-attribute graphs; for each of these, one numeric vector is generated by using state-of-the-art graph embedding techniques; (ii) the obtained vectors are concatenated in one representative vector using a multi-view learning integration technique; (iii) the size of such a vector is reduced through deep autoencoders.Experiments have been conducted on the music style recognition problem. We focus on the corpus of 4-voice J. S. Bach’ compositions. First, such a corpus has been decomposed and translated into graph-based structures corresponding to the music scores. Then, the proposed method is applied to generate the embedding vectors from the obtained graphs. Finally, a Random Forest model trained on such obtained vectors is used for generating novels music compositions in the learned style. Results obtained show the effectiveness of the proposed approach.