融合“科学-技术”知识关联的高颠覆性专利预测方法

Integrating Science-Technology Knowledge Linkage to Predict Disruptive Patents

颠覆性技术的识别与预测研究在服务国家重大科技战略发展需求、保障国家科技产业安全等方面具有重要意义。本文将专利家族视为技术单元,从对技术知识空间的改变视角定义高颠覆性专利,基于世界专利统计数据库(Worldwide Patent Statistical Database,PATSTAT)和微软学术论文(Microsoft Academic Graph,MAG)数据库,对专利的颠覆性、技术特征及“科学-技术”知识关联特征进行测度分析,并在此基础上提出了融合“科学-技术”知识关联的高颠覆性专利预测方法。本文将高颠覆性专利预测问题转化为监督式二分类任务,给定专利在其公开当年的“科学-技术”知识关联和其他技术特征,以其5年后的颠覆性指标值高低作为预测目标,训练机器学习模型。研究结果表明,①高颠覆性专利具有前置知识少且非主流、技术团队实力强、商业价值被低估、长期影响力大的特点;②专利的“科学-技术”知识关联属性是对其颠覆性进行预测的重要特征;③LightGBM(light gradient boosting machine)模型在综合性能与训练效率上取得了最佳表现,在半导体器件与电数字数据处理领域的实证结果验证了模型的有效性。但颠覆性技术的预测仍是一个困难的任务,后续研究可尝试从专利语义特征与结合多源数据等角度进一步提升性能表现。

Identifying and predicting disruptive technologies is critical to the national need for strategic development.This study treats patent families as technology units and calculates their disruption index,technological features,and Science-Technology(S&T)association features based on two large-scale patent(PATSTAT)and bibliographic(MAG)datasets.An approach to predicting potential disruptive patents is proposed.The prediction is considered a supervised binary classification task,which predicts the patent disruptiveness in five years,given the features calculated in the year it was published.Our results show that:(1)disruptive patents are characterized by less prior knowledge,stronger teams,an underestimated commercial value,and a higher long-term impact;(2)S&T linkage is an important feature in predicting disruptive patents;and(3)the LightGBM model achieves the best results in terms of performance and efficiency.However,the prediction of disruptive patents remains difficult.Future studies should consider incorporating semantic features and multiple data sources to improve the performance.