A Comparison of Multi-label Instance Selection Methods Using Problem Transformation Approach for Music Emotion Recognition
Miao Xu & Jae-Sung Lee Graduate Student, Kyungpook University & Associate Professor, Chungang University Affective engineering refers to the development of technologies that can analyze human emotions and applying such technologies to environmental design for promoting convenience and comfort in human lives. Numerous studies in affective engineering have been conducted on the quantification of classical literature, music, and art works that evoke and transit nonverbal elements. Music emotion recognition involves the identification of music emotions expressed or inherently present in the music based on quantified data and incorporating them into various music services, thus enhancing the user's convenience. In conventional music emotion recognition, music data is generally created by extracting information from a music segment for mitigating efficiency issues. If the sampled music segment is unrepresentative of the entire music, it can act as noise in the learning process and hence should be eliminated before initiating the learning process. This study proposes various multi-label instance selection algorithms and compares their performances with the help of thirteen multi-label datasets. Obtained experimental results demonstrate that the instance selection method with label powerset transformation can achieve the best performance. Key words： Music Emotion Recognition, Multi-label Learning, Multi-label Instance Selection | 
