Abstract

The influence of music on human emotions and behavior has been recognized for a long time, with applications spanning therapeutic interventions and veryday mood regulation. This Master thesis explores the potential of music to modulate emotional states in real-world settings by developing a context-aware system for playlist generation aimed at enhancing specific mood parameters. This study employs a two-stage approach to model training. First, a representation model is pre-trained to learn multimodal music representations by capturing similarities across various modalities, including Mel spectrograms, genre, tags, emotional tags, and lyrics. This pre-training enables the model to learn to represent a music track in multiple ways over multiple modalities, finding representations that maximize the mutual information between the Mel spectrogam and another modality. Subsequently, a mood context representation model is fine-tuned to align music track representations with the contextual features of listening sessions. Leveraging these learned representations, the system predicts the emotional and contextual suitability of tracks and their impact on mood, thereby facilitating the creation of playlists tailored to evoke specific emotional effects and suit diverse situational contexts. This Master thesis systematically evaluates the proposed models and compares them against state-of-the-art representation models. While the proposed model does not exhibit superior generalization across multiple benchmark datasets, it demonstrates improved similarity between representations of tracks grouped by emotional tags and emotional numeric information. To assess the system’s effectiveness, two user studies were conducted. The first study collected music listening sessions and contextual information to enhance model training, while the second examined user perceptions of the generated playlists concerning mood regulation and contextual relevance. The findings underscore the inherent subjectivity of music perception, as the same playlist elicited varied responses among different users. The studies further emphasize the importance of user preferences, revealing that ratings fluctuate based on context and are highly individualized. This Master thesis contributes to the field of music recommendation by demonstrating the feasibility of integrating contextual and emotional data into automatic playlist generation. Additionally, it provides insights into the strengths and limitations of current representation models and the role of user preferences in music suitability for certain contexts.

Citation

BibTeX

@misc{Hausberger2025master-thesis,
    title = {Generating Context-based Music Playlists for Mood Regulation: An Approach Based on Contrastive Learning},
    author = {Hausberger, Anna},
    school = {Johannes Kepler University Linz},
    year = {2025}
}