Computational study demonstrates accurate entity classification and link prediction across innovation ecosystems, indicating the power of heterogeneous graph modeling.
Key Points
Develop a heterogeneous graph neural network framework to model multi-entity stakeholders and predict relational dynamics within innovation and entrepreneurship ecosystems.
Constructed a heterogeneous graph linking enterprises, universities, and governments via technological cooperation, investment, and policy support edges.
Encoded structural and behavioral features into multimodal vectors and applied a heterogeneous graph attention network to generate low-dimensional node embeddings.
Evaluated ecosystem subgroup partitioning, core node identification, and link prediction performance using classification accuracy and area under the curve metrics.
Achieved node classification accuracies of 0.92 for enterprise, 0.91 for university, and 0.95 for government nodes.
Yielded link prediction AUC values exceeding 0.92 across technological cooperation, investment, and policy support relations.
Confirmed network structure preservation and entity differentiation through node centrality correlations and semantic separation analyses.