Music Emotion Classification Based on Heterogeneous Graph Neural Networks

The classification of musical emotions is crucial for the indexing, structuring, searching, and recommending of tracks and albums across various music platforms. Consequently, the automated categorization of musical emotions has become a vital element in nearly all music applications. Recent studies have mainly concentrated on utilizing textual, audio, or multimodal data for genre classification, frequently neglecting the impact of singers, composers, and listener preferences. In practice, composers possess unique compositional styles, listeners have varied musical preferences, and singers focus on particular music genres. These different viewpoints offer significant insights into the classification of musical emotions, greatly enhancing the effectiveness of classification performance. In this paper, we introduce a novel heterogeneous graph neural network (HGN) that models the relationships of music emotion preferences among singers, composers, and listeners, in order to generate accurate node feature representations for downstream tasks. The experimental results show that our model significantly outperforms current state-of-the-art (SOTA) methods on two datasets for music emotion classification.