Predicting Musical Genres using Deep Learning and Ensembling
Automatic Music Genre Classification is a core problem in the Music Information Retrieval space. The classification approach detailed in this paper involves: using musical features from the Million Song Dataset, augmenting the musical dataset with lyrics and cover art images, building a deep learning model for each of the three different types of inputs, and then ensembling the predictions from the individual models using a gradient boosted machine. Ensembling resulted in an 8.6% increase in F1 score over the best individual model while maintaining a similar level of accuracy. This framework may be successfully applied to other problems with multimodal inputs.