An Investigation of Different Positive and Negative Training Policies for the Task of Hierarchical Music Genre Classification

Within the growth of the music stream companies, the need to classify correctly the musics in their catalogs by genres becomes a really important task. With that, they can suggest songs suitable to the music taste of the user, offering a better user experience. Also, because of the size of their catalogs, this task ought to be done automatically. One way to handle the music genre classification problem is to employ a hierarchical classification approach. In this approach, there is a parent-child relation between the genres, characterizing a hierarchical classification problem. This kind of problem can be tackled by using different approaches, in this paper we investigated the use of the Local Classifier per node approach and different positive and negative training policies. The main contribution of this work is to investigate the impact of different positive and negative training samples policies to train the different classifiers with different versions of the Free Music Archive Database. The computational results show that extracting features from just a part of the music produce similar results from analysing the entire music. In addition, the Random Forest was the best classifier and the less inclusive policy produced the highest scores.