Multi-Label Playlist Classification Using Convolutional Neural Network
Guanhua Wang, Chia-Hao Chung, Yi‐An Chen, Homer H. Chen · 2018
With the popularization of music streaming services, millions of songs can be accessed easily. The effectiveness of music access can be further enhanced by making the labels of a music playlist more indicative of the theme of the playlist. However, manually classifying playlists is laborious and often requires domain knowledge. In this paper, we propose a novel multi-label model for playlist classification based on a convolutional neural network. The network is trained in an end-to-end manner to jointly learn song embedding and convolutional filters without the need of feature extraction from audio signals. Specifically, the song embedding vectors are concatenated as a matrix to represent a playlist, and the convolutional filters for playlist classification are applied to the playlist matrix. We also propose two augmentation techniques to prevent over-fitting of playlist data and to improve the training of the proposed model. Experimental results show that the proposed model performs significantly better than the support vector machine and k-nearest neighbors models.