基于时序图神经网络的潜在高价值专利识别研究  

Potential High-Value Patent Identify Based on a Time-Series Graph Neural Network

作　　者：周潇 王博 胡玉琳 韦楚楚 Zhou Xiao;Wang Bo;Hu Yulin;Wei Chuchu(School of Economics and Management,Xidian University,Xi’an 710126)

机构地区：[1]西安电子科技大学经济与管理学院,西安710126

出　　处：《情报学报》2024年第6期697-711,共15页Journal of the China Society for Scientific and Technical Information

基　　金：国家自然科学基金面上项目“重组创新视角下新兴共性技术识别及突破路径预测研究”(72374165)。

摘　　要：高价值专利是构建当前“国内国际双循环”新发展格局的核心资源,也是促使我国在国际经济新秩序中立足战略制高点、全面推进科技自立自强的核心要素,准确识别潜在的高价值专利是对其进行价值培育与技术转化的关键性步骤。本文在充分挖掘中国专利奖获奖专利特征的基础上,综合利用Patent-BERT(bidirectional encoder representations from transformers for patent)与图深度学习算法,在融合专利评估指标、文本特征的基础之上,提出了基于图卷积神经网络(graph convolutional network,GCN)与长短时记忆网络(long short-term memory,LSTM)的潜在高价值专利识别模型。本文的创新点主要体现在两个方面:(1)修正了已有研究中仅关注诸如专利增长速度、合作潜力等“数量”特征而缺乏对文本语义深度理解的弊端,从文本语义与专利计量维度构建专利价值的表示模型;(2)考虑到专利价值的时序变化性,从动态视角探索了专利价值的演化规律,为专利价值的挖掘与评估提供了新的研究思路。最后,本文对node2vec、doc2vec、GCN、MLP(multilayer perceptron)等多种模型进行性能对比,研究结果表明,本文模型在多项指标上的表现均优于对照模型,从而有效验证了本文方案的高效性与稳健性。High-value patents are primary resources in constructing the current“dual circulation”development pattern at both domestic and international levels.They also play a pivotal role in positioning China at a strategic high ground in the new international economic order and comprehensively advancing technological self-reliance and self-strengthening.Precisely identifying potential high-value patents is a crucial step for nurturing their value and promoting technological transfer.Based on an in-depth analysis of the characteristics of patents that have won the China Patent Award,this study combines the use of Patent-BERT(bidirectional encoder representations from transformers for patent)and graph deep learning algorithms.By integrating patent evaluation indicators and textual features,we propose a potential high-value patent identification model based on graph convolutional networks(GCNs)and long short-term memory(LSTM)networks.The two main innovative aspects of this research are as follows:(1)Addressing the shortcomings of previous studies that only focused on“quantitative”features such as patent growth rate and collaboration potential and lacked deep semantic understanding of the text.We build a patent value representation model from both textual semantics and patent metrics dimen‐sions.(2)Considering the temporal variability of patent value,we explore the evolutionary rules of patent value from a dy‐namic perspective,providing a new research approach for patent value mining and assessment.Finally,we compare the performance of various models,including node2vec,doc2vec,GCN,and multilayer perceptron(MLP).The results indicate that our model outperforms the control models across multiple indicators,thereby effectively validating the efficiency and robustness of our research approach.

关 键 词：战略情报预判 高价值专利识别 多源特征融合 时序图神经网络 表示学习 

分 类 号：G306[文化科学]