Combined Transfer and Active Learning for High Accuracy Music Genre Classification Method

Abstract

Music genre classification system has been widely used by commercial music apps or professional music systems. At the same time, the growing complexity of genres and the rapped combination of multiple genres continuously challenge the accuracy of classical algorithms. Therefore, the need for more advanced genre classification system has been triggered. This paper propose a novel active transfer music genre classification method (ATMGCM) for musical genre classification. After comparing the ATMGCM with SVM and Random Forest, the ATMGCM algorithm has higher accuracy in massive database or with noises. In addition, the simulation results indicate that the ATMGCM has a significant decrease in niche genres. Therefore, the Transfer learning is applied to transfer knowledge from the source dataset to the target dataset, and active learning is applied to determine informative labels of a small part of samples from unlabeled datasets. Through experiments, this study shows that ATMGCM algorithm is effective at classifying genres. Furthermore, this method only need to label about 10%-15% of unlabeled data and can still achieve significant performance improvement. In the future, the research will use RNN model (LSTM) to improve the performance of classification results.