Vocal Learner Ability Profiles Integrating Graph Neural Networks

Existing competency assessments often evaluate individual skills independently and fail to represent the structural dependencies between professional knowledge and practical performance. This paper applies Graph Neural Networks to construct dynamic ability profiles for vocal learners based on knowledge-structure modeling. Practice records and curriculum data are integrated into a heterogeneous graph containing learners, skills, and knowledge points. An R-GCN first generates initial embeddings through relation-specific message passing, preserving the semantic effects of different edge types. A Temporal GAT then applies temporal encoding to dynamic graph sequences to capture the evolution of learner abilities over time. The final time-series ability embedding is mapped to five dimensions, including pitch, rhythm, breath, resonance, and emotional expression. Experimental results show that structural modeling achieves average node reconstruction accuracy of at least 0.81 and relationship prediction F1 of at least 0.80. A dynamic embedding similarity slope of 0.82 and ability-change direction cosine of 0.85 indicate high temporal consistency in tracking ability evolution. The proposed framework supports structured inference of learner competence and provides a technical basis for personalized vocal-training intervention.