Abstract: Technology convergence is essential to technological innovation, so it is important to achieve its early prediction. Current research on technology convergence prediction is mostly based on 4-bit IPC co-occurrence prediction, but technology convergence often occurs at a more micro-level and needs to be based on finer-grained data and methods. However, the problem of sparse feature dimensions and sample imbalance in fine-grained technical convergence prediction has hindered its further development. To address this problem, this paper proposes a fine-grained technical convergence prediction method based on graph neural networks. A graph neural network model GraphSAGE, able to integrate multi-dimensional features, is introduced, and the sampling strategy is improved to achieve fine-grained technical convergence prediction. The research results show that this method is feasible for fine-grained technical convergence prediction as it improves the AUC value by 10% and 4.1% respectively compared to the baseline model using traditional similarity metric features and the traditional graph representation learning model.

Key words: Graph neural networks, GraphSAGE, Technology convergence prediction, Link prediction, Patent data