Multi-Channel Graph Convolutional Network for Evaluating Innovation Capability Toward Sustainable Seed Enterprises

Abstract

The innovation capability of seed enterprises reflects their core competitiveness and serves as a vital foundation for sustainable agricultural development and modernization. Therefore, evaluating this capability is of great importance. However, existing evaluation methods primarily focus on internal enterprise attributes, overlooking the complex inter-enterprise relationships and lacking sufficient feature fusion capabilities to capture latent information. To address these limitations, this paper proposes a Multi-Channel Graph Convolutional Network (MGCN) model that integrates enterprise attributes with three types of relational graphs. The model adopts a multi-channel architecture for feature extraction and employs a gated attention mechanism for cross-graph feature fusion, jointly considering node features and relation information to improve prediction accuracy. Experimental results demonstrate that MGCN achieves an average accuracy of 83.59% under five-fold cross-validation, outperforming several mainstream models such as Random Forest and traditional GCN. Case studies further reveal that MGCN not only captures key features of individual enterprises but also leverages features and label distribution from neighboring enterprises, facilitating more context-aware classification decisions. In conclusion, the MGCN model provides an effective method for the intelligent evaluation of innovation capability in seed enterprises and supports the formulation of sustainable strategic plans at both the national and enterprise level.