Knowledge graph-enhanced heterogeneous graph neural network for scientific talent innovation potential identification

Area of Science:

• Computer Science
• Bibliometrics
• Research Management

Background:

• Traditional talent evaluation relies on static bibliometrics, missing dynamic research evolution and innovation potential.
• Existing methods struggle with the complexity and multi-faceted nature of scientific talent.

Purpose of the Study:

• To propose a novel knowledge graph-enhanced heterogeneous graph neural network framework for identifying innovation potential in scientific talents.
• To develop a more dynamic and comprehensive approach to talent evaluation.

Main Methods:

• Constructed a comprehensive knowledge graph using multi-source heterogeneous academic data (researchers, publications, institutions, topics).
• Employed meta-path-based attention mechanisms for selective information aggregation.
• Utilized a gated fusion strategy to combine knowledge graph embeddings and academic network features for talent representation learning.

Main Results:

• Achieved 85.21% accuracy and 0.9014 AUC-ROC score on a dataset of 128,456 researchers.
• Demonstrated a 6.3% improvement over state-of-the-art baseline models.
• Showed particular effectiveness in identifying early-career researchers with high innovation potential.

Conclusions:

• The proposed framework offers a generalizable methodology for knowledge-augmented graph representation learning.
• Provides practical solutions for intelligent talent management and addresses cold-start problems in conventional evaluation systems.