Instance Selection for Music Genre Classification using Heterogeneous Networks

In scenarios involving musical data, there are usually high-dimensional data and different modalities, such as audio and text, that cost more in machine learning tasks. Instance selection is a promising approach as pre-processing step to reduce these challenges. With the intent to explore the multimodality in music information, we introduce musical data instance selection into heterogeneous network models. We propose and evaluate ten different heterogeneous networks to identify more representative relationships with various musical features related, including songs, artists, genres, and melspectrogram. The results obtained allow us to define which network structure is more appropriate considering the volume of available data and the type of information that the features have. Finally, we analyze the relevance of the musical features, and the relationship does not contribute for instance selection.