Induced Emotion-Based Music Recommendation through Reinforcement Learning

Music is widely used for mood and emotion regulation in our daily life. As a result, many research works on music information retrieval and affective human-computer interaction have been proposed to model the relationships between emotion and music. However, most of these works focus on applications in a context-sensitive recommendation that considers the listener’s emotional state, but few results have been obtained in studying systems for inducing future emotional states. This paper proposes Moodify, a novel music recommendation system based on reinforcement learning (RL) capable of inducing emotions in the user to support the interaction process in several usage scenarios (e.g., games, movies, smart spaces). Given a target emotional state, and starting from the assumption that an emotional state is entirely determined by a sequence of recently played music tracks, the proposed RL method is designed to learn how to select the list of music pieces that better “match” the target emotional state. Differently from previous works in the literature, the system is conceived to induce an emotional state starting from a current emotion instead of capturing the current emotion and suggesting certain songs that are thought to be suitable for that mood. We have deployed Moodify as a prototype web application, named MoodifyWeb. Finally, we enrolled 40 people to experiment MoodifyWeb, employing one million music playlists from the Spotify platform. This preliminary evaluation study aimed to analyze MoodifyWeb’s effectiveness and overall user satisfaction. The results showed a highly rated user satisfaction, system responsiveness, and appropriateness of the recommendation (up to 4.30, 4.45, and 4.75 on a 5-point Likert, respectively) and that such recommendations were better than they thought before using MoodifyWeb (6.45 on a 7-point Likert).