Abstract Traditional talent evaluation methods predominantly rely on static bibliometric indicators that fail to capture the dynamic evolution patterns and potential innovative capabilities of researchers. This study proposes a novel knowledge graph-enhanced heterogeneous graph neural network framework for identifying innovation potential in scientific talents. The framework integrates multi-source heterogeneous academic data to construct a comprehensive knowledge graph encompassing researchers, publications, institutions, and research topics, while employing meta-path-based attention mechanisms to selectively aggregate information from diverse entity types and relationships. A gated fusion strategy adaptively combines semantic embeddings from knowledge graphs with structural features from academic networks, enabling comprehensive talent representation learning. Experimental validation on a dataset containing 128,456 researchers across multiple disciplines demonstrates superior performance, achieving 85.21% accuracy and 0.9014 AUC-ROC score, representing significant improvements of 6.3% over state-of-the-art baseline models. The proposed approach exhibits particular effectiveness in identifying early-career researchers with high innovation potential, addressing cold-start problems inherent in conventional evaluation systems. This research provides a generalizable methodology for knowledge-augmented graph representation learning and offers practical solutions for intelligent talent management in research institutions.
Rong Wang (Mon,) studied this question.