Technology convergence prediction based on temporal heterogeneous graph neural networks

• Propose a temporal HGNN-based model using structural & semantic features for TOA. • Analyze TC opportunities from multiple levels with the patent classification system. • Experiments on the large language model technology convergence prediction show the superiority of the proposed model. Most studies on technology convergence prediction typically focus on the co-occurrence relationship of technology elements in patent data, neglecting the joint effects of internal and external factors on technology convergence. This study constructs a novel temporal heterogeneous graph neural networks-based technology convergence prediction framework considering structural and semantic features (THGNN-TCP) to dynamically identify technology opportunities at multiple granularities. The framework takes into account the co-occurrence relationship, semantic distance and knowledge flow among technologies, as well as the linkage between science and technology to capture the changes in the internal and external influencing factors of technology. Specifically, we introduce a node type-aware strategy and a talking-heads attention mechanism to capture the structural features among technologies, while leveraging LSTM to discern the temporal dependencies inherent in these technological interactions. Our experiments on datasets in the field of large language models demonstrate that THGNN-TCP significantly outperforms a series of baseline approaches in terms of AUC and F1 metrics, and is capable of identifying potential technological convergence opportunities from a multi-granularity perspective.