基于图神经网络的细粒度技术会聚预测方法研究

A Fine-grained Technology Convergence Prediction Method Based on Graph Neural Networks

技术会聚作为技术创新的一种重要实现途径,实现其早期预测具有重要意义。当前技术会聚预测多以4位专利分类号的粗粒度实体作为分析对象,然而,技术会聚往往发生在更微观的层次,需要以更细粒度的数据与方法为基础。通过细致比对分析,本研究发现,细粒度技术会聚预测存在特征维度稀疏、正负样本不均衡问题,阻碍了相关研究方法的推广。对此,本研究提出基于图神经网络的细粒度技术会聚预测方法,通过引入能够综合使用多维度特征的图神经网络模型GraphSAGE,并改进样本采样策略,从而确保技术会聚预测能够在细粒度层次上开展。研究结果表明,本研究方法相较于基于传统相似性度量特征的基线模型和仅关注于网络结构的传统图表示学习模型,AUC值能够分别提升10%和4.1%,是开展细粒度技术会聚预测的一种可行方法。

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.