Graph Embedded Multiple Kernel Extreme Learning Machine for Music Emotion Classification

Xixian Zhang, Zhijing Yang, Jinchang Ren, Meilin Wang, Wing-Kuen Ling

Abstract: Music emotion classification is one of the most importance parts of music information retrieval (MIR) because of its potential commercial value and cultural value. However, music emotion classification is still a tough challenge, due to the low representation of music features. In this paper, a novel Extreme Learning Machine (ELM), combining graph regularization term and multiple kernel, is proposed to enhance the accuracy of music emotion classification. We use nonnegative matrix factorization (NMF) to find the optimal weights of combining multiple kernels. Furthermore, the graph regularization term is added to increase the relevance between predictions from the same class. The proposed Graph embedded Multiple Kernel Extreme Learning Machine (GMK-ELM) is tested on three music emotion datasets. Experiment results show that the proposed GMK-ELM outperforms several well-known ELM methods.