Personalized Emotion-based Music Classification With Reinforcement Learning

Music is an important carrier of emotion. People can’t live without music. With the rapid growth in the number of digital music, the need for music emotion classification and retrieval is also increasing. Current research on music emotion classification usually takes it as a static problem and there’s little concern about users’ subjective effect on the understanding of music emotions.We studied personalized emotion-based music classification system with reinforcement learning. We used an emotion space containing 12 emotion categories based on the classical psychology emotion model VA model, extracted the audio features from the songs, trained a brand new music emotion classification model of high accuracy by applying and comparing several classification algorithms. Taking into account the classification error of the model and the user’s subjective understanding of music emotions, we analyzed user behavior during music playing and built a novel dynamic model based on reinforcement learning. This model can dynamically adjust the emotion classification results by learning user behavior, which helps personalize user’s emotion tastes and preferences of music.We collected 600 songs from commercial websites covering 12 emotion types, applied audio feature extraction and feature selection for the songs, built a new training set of music emotion classification problem and implemented a novel prototype system. We measured the accuracy of the static music emotion classification model and verified the effectiveness of the dynamic learning model through experiments. The result of a small-scale user study shows that our prototype system works well.