In real-world problems, modeled as machine learning tasks, the datasets are typically unbalanced, meaning that some classes have much more instances than others. In the Music Information Retrieval field it is not different and songs datasets usually are very unbalanced. Considering this scenario, we propose a novel approach to face the class imbalance problem applied to music genre classification. The proposed method uses vertical sliced spectrograms extracted from the songs' audio signal to apply oversampling and undersampling into the minority and majority classes, respectively. The experimental results for F-Score measure showed that our approach was able to beat the best result of Random Undersampling technique by 0.086, using MultiLayer Perceptrons. Besides, comparing to the baseline results, our approach significantly increased the individual results for all the minority classes.

Imbalanced classification tasks have been studied by the research community for a long time. Numerous problems have been identified with standard approaches and new proposals have been put forward for addressing these relevant tasks. Surprisingly, the same attention has not been given to predictive tasks with a numeric target variable, i.e. regression. However, similar problems occur on these domains, when the target of the end-user is the performance on a subset of rare values of the target variable. As in classification standard evaluation metrics fail, and new approaches are required to bias the learning algorithms to the end-user goals.