Multi-source music knowledge graph construction and prosodic evolution trajectory prediction via DRL-VAE...

Area of Science:

• Computational Musicology
• Artificial Intelligence
• Data Science

Background:

• Heterogeneous music data across modalities poses challenges for unified retrieval and analysis.
• Existing methods struggle with integrating diverse acoustic, symbolic, and semantic music information.

Purpose of the Study:

• To propose a novel framework for multi-source music knowledge graph construction and prosodic evolution trajectory prediction.
• To enable systematic entity alignment across disparate music data sources.
• To bridge music knowledge representation and computational prosodic analysis.

Main Methods:

• Integration of deep reinforcement learning (DRL) with variational autoencoders (VAE).
• Hierarchical ontological schema for accommodating acoustic, symbolic, and semantic modalities.
• DRL-VAE collaborative mechanism for latent representations and knowledge graph completion.
• Prosodic trajectory prediction using generative modeling, long-horizon planning, and graph neural networks.

Main Results:

• Achieved a 0.456 MRR on link prediction, a 17.8% improvement over baselines.
• Trajectory prediction attained an F1-score of 0.867 with superior generalization.
• Demonstrated effectiveness on large-scale public music datasets.

Conclusions:

• The proposed DRL-VAE framework effectively constructs music knowledge graphs and predicts prosodic evolution.
• This integrated methodology enhances music information retrieval and computational prosodic analysis.
• The approach shows strong performance and generalization across diverse musical styles.