Personalized Emotion-based Multi-label Classification For Music

Online music services have experienced rapid growth in recent years. They have provided convenient platforms for users to get music. For the convenience of users’ selection, online music platforms generally classify music based on some categories. Emotion-based music classification is a common music classification category. One piece of music may involve more than one different emotions, and single-label classification cannot cover this multiplicity, so we need multi-label classification to model emotion category of a piece of music. Furthermore, music emotion perception is by nature subjective, therefore providing a personalized music emotion classification is necessary.This thesis studies the problem of personalized emotion-based multi-label classification for music, and it contains two parts:obtaining of ground truth data of music emotion and personalized emotion classification. To deal with these problems, first, we propose an approach to compute ground truth data by using social information and emotion tags provided by users. It can reduce users’ participation. Next, we propose an approach to map ground truth data of music emotion to specific multi-label emotion categories. Finally, we use convolutional neural network and random k-labelsets method to classify music based on emotion.About experiments, we have collected a lot of real data from online music platform as the data source of music emotion classification. Through extensive experiments, the results of experiments demonstrate the effectiveness of our proposed methods.