Predicting technological convergence with multi-channel graph neural networks: A case study of CRISPR

Abstract

Identifying emerging technological convergence is essential for anticipating future innovation trajectories. Existing approaches typically rely on either International Patent Classification (IPC) co-occurrence networks, which capture general combination frequencies, or association rule networks, which emphasize statistically significant and often higher-order relationships. However, these two structural views are rarely integrated, limiting their effectiveness in representing both the breadth and depth of technological linkages. To address this gap, we propose a Multi-Channel Graph Convolutional Network (MC-GCN) that treats IPC co-occurrence and association rule networks as structurally distinct inputs. While co-occurrence data reflect raw interaction patterns, association rules—derived via data mining—serve as a refined signal that highlights meaningful and potentially multi-IPC convergence patterns. Our model encodes each view through separate GCN channels and fuses their embeddings within a unified representation space. To establish a comprehensive evaluation, we also include topological link prediction baselines such as Common Neighbors, Adamic–Adar, and Preferential Attachment in our comparative analysis. Applied to CRISPR-related patent data, the MC-GCN significantly outperforms single-channel models, achieving an AUC of 0.973 when combined with XGBoost. Furthermore, five predicted IPC combinations were validated in newly granted patents in early 2025, demonstrating the model’s practical utility in forecasting real-world technological convergence.