Emotion-based Multimodal Music Classifier for Recommender Systems

In recent years, advancements in artificial intelligence have driven a growing demand for personalized user experiences across various digital platforms. In the music domain, this trend is reflected in the need for more sophisticated recommendation systems beyond traditional collaborative filtering methods. This thesis introduces an emotion-based multimodal music classifier, leveraging both audio features and song lyrics to capture the emotional content of music. By focusing on song content and emotional attributes, this approach aims to lay the groundwork for recommendation systems capable of providing users with a more customized and emotionally resonant experience, also addressing the cold-start problem typical of collaborative filtering-based recommendation. Following an overview of existing literature on the topic and the examination of the challenges posed by the specific field of interest, the first contribution of this work is the creation of a suitable dataset for Music Emotion Recognition: this is achieved by extending a subset of the Music4All-Onion dataset with emotion-based labels for song lyrics using an eight-class emotional model. Audio data, available in the form of pre-extracted acoustic features, is analyzed using unsupervised machine learning methods to meaningfully model the underlying structures and patterns associated with emotional music content. Different baselines methods are identified as benchmarks to comparatively evaluate the proposed approaches. The final emotion-based multimodal classifier primarily relies on textual data in the form of song lyrics, incorporating acoustic information as an auxiliary feature to enhance the emotional classification. The model achieves promising results in the context of Music Emotion Recognition, considering the data availability issues characterizing research in this field.