Ontology-Driven Hierarchical Learning for Raga Identification

Ragas are the fundamental melodic frameworks in Indian classical music, each evoking distinct moods and traditionally associated with specific times of the day for performance. These time associations not only guide when a Raga should be sung but also enhance its aesthetic and emotive qualities, resonating with the environment and the mood it seeks to convey. While Raga identification has been largely approached as a standalone classification task, we explore whether incorporating the time-of-day information, an essential aspect of Indian classical music, can enhance the accuracy of Raga identification and contribute to a more robust model. To investigate this, we propose an ontology-based approach that integrates hierarchical learning with constraints for the task. These constraints enforce consistency between predictions of child classes (Ragas) and their parent classes (singing times), ensuring that the model learns a structured representation where the associated child classes of a particular parent class are closely aligned in the latent space. Experimental results demonstrate that incorporating these constraints leads to significantly better performance compared to models without constraints. The approach highlights the importance of using hierarchical relationships for Indian Art Music analysis.