基于图神经网络和粒子群算法的技术预测模型

Graph Neural Network-based and Particle Swarm Optimization Technological Prediction Model

制定科技产业政策需要科学预测技术发展趋势,专利申请书能够细粒度地表述技术特征,对专利申请书进行分析并构建技术预测模型有助于提高技术预测准确率。本文以专利申请书为研究对象,构建一种基于粒子群算法和图神经网络的技术预测模型,以人工智能领域中美国公司在华布局的594项专利申请书作为研究对象,在技术预测任务上以指数平滑法、平均移动法、SVR(support vector regression)、GRU(gate recurrent unit)、RNN(recurrent neural network)等算法为基线进行实验。结果表明,本文模型的准确性优于现有技术预测模型,可揭示技术新特征的形成过程。采用此模型对人工智能领域美国在华专利布局进行预测,并对重点领域的具体技术特征进行分析,得出了当前美国在华技术布局的趋势、具体特征与空白点。该模型能够提升技术预测准确性,可为科技产业布局与科研人力物力投入提供决策依据。

Effective prediction of technological development trends is crucial for policymaking in many technological industries.As patent application forms express technical features in an in-depth manner,they can be used to train a model to improve the technology prediction.This study constructs a technology prediction model on the basis of particle swarm optimization(PSO)and graph neural networks to help improve the prediction accuracy of the future development trend of a technology field and the characteristics of its emerging technologies.Using 594 artificial intelligence patent applications arranged in China by US companies over the last two decades as the research objects,this study conducted an experiment and found that the suggested model obtains higher accuracy than baseline algorithms,such as the exponential smoothing method,moving average method,support vector regression,gate recurrent unit,and recurrent neural network.Moreover,the model can reveal the formation process of the new characteristics of the technologies.The study predicts the layout of US patents in China to investigate the trend,characteristics,and gaps in the current US technology layout in China.The proposed graph neural network-based and PSO technological prediction model can improve the technology prediction accuracy and support the decision making in the layout of technological industries and the funding of scientific research.