Machine Learning Approaches for Mood Classification of Songs toward Music Search Engine

Abstract

Human often wants to listen to music that fits best his current emotion. A grasp of emotions in songs might be a great help for us to effectively discover music. In this paper, we aimed at automatically classifying moods of songs based on lyrics and metadata, and proposed several methods for supervised learning of classifiers. In future, we plan to use automatically identified moods of songs as metadata in our music search engine. Mood categories in a famous contest about Audio Music Mood Classification (MIREX 2007) are applied for our system. The training data is collected from a LiveJournal blog site in which each blog entry is tagged with a mood and a song. Then three kinds of machine learning algorithms are applied for training classifiers: SVM, Naive Bayes and Graph-based methods. The experiments showed that artist, sentiment words, putting more weight for words in chorus and title parts are effective for mood classification. Graph-based method promises a good improvement if we have rich relationship information among songs.