摘要
融合异构信息进行专利交易推荐可以促进交易,但存在因忽略专利属性而影响推荐结果的问题。本研究提出基于属性异构网络(attribute heterogeneous network,AHN)表示学习的专利交易推荐模型(patent transaction recommendation based on AHN representation learning,AHNRL-PTR)。首先筛选专利和组织中影响专利交易的属性;其次构建专利交易AHN,然后在AHN中引入网络表示学习,并基于多维高斯分布解决节点表示的不确定性,基于KL散度(Kullback-Leibler divergence)解决节点间距离非对称性。最后,以粤港澳大湾区有效发明授权专利数据进行实证研究,得出结论:第一,相比于metapath2vec、TADW(text-associated DeepWalk)和AHNRL-PTR模型的两个变体方法,AHNRL-PTR模型的推荐精度最高,超过86%,说明融合组织及专利属性,并聚焦节点表示的不确定性和非对称性问题的解决,能大幅提高推荐精度;第二,在非准确指标IntraSim和Popularity上,AHNRL-PTR的表现优于metapath2vec和两个变体方法,反映该方法的推荐结果具有一定的多样性,且可以挖掘推荐冷门专利;第三,基于两个非准确指标将组织聚类为六类,分别为中介型、领域骨干型、研究型、族群型、成长型、专业型,体现了推荐结果的可解释性和个性化水平。本研究可为专利交易智能化推荐服务提供决策支持。
Patent transaction recommendation is an important means of fusing heterogeneous information to facilitate transactions;however,the recommendation results are often affected by the disregard of patent attributes,which represents a research problem.This study proposes a patent transaction recommendation model based on attribute heterogeneous network(AHN)representation learning(AHNRL-PTR),which firstly filters the patent and organizational attributes affecting patent transaction,secondly constructs a patent transaction AHN,then introduces network representation learning in AHN,and finally,uses multidimensional Gaussian distribution and Kullback-Leibler divergence to solve the problems of node representation uncertainty and distance asymmetry between nodes.Finally,an empirical study with the valid invention granted patent data in the Greater Bay Area concluded that:first,compared to the metapath2vec,text-associated DeepWalk(TADW),and variant methods of the AHNRL-PTR model,the AHNRL-PTR model has the highest recommendation accuracy(more than 86%),indicating that fusing organizational and patent attributes and focusing on the solution of the uncertainty and asymmetry problem of node representation can substantially improve recommendation accuracy;second,the values of the non-accurate metrics IntraSim and Popularity of AHNRL-PTR are smaller than those of metapath2vec,AHNvec-PTR,and AHNsy-PTR methods,reflecting the diversity of this method’s recommendation results and its advantage in recommending niche cold patents;third,the organizations are clustered into the following six categories based on IntraSim and Popularity:intermediary,domain backbone,research,community,growth,and professional,which reflect the recommendation results’interpretability and personalization level.Given the results,this study provides decision support for intelligent recommendation services for patent transactions.
作者
何喜军
吴爽爽
武玉英
才久然
庞婷
Chee Seng Chan
He Xijun;Wu Shuangshuang;Wu Yuying;Cai Jiuran;Pang Ting;Chee SengChan(School of Economics and Management,Beijing University of Technology,Beijing 100124;Department of Artificial Intelligence,Faculty of Computer Science and Information Technology,University of Malaya,Kuala Lumpur 50603;Network and Information Center,Xinxiang Medical University,Xinxiang 453000)
出处
《情报学报》
CSSCI
CSCD
北大核心
2022年第11期1214-1228,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“异构信息网络下技术供需匹配模型与对接路径研究”(71974009)
国家自然科学基金项目“工程教育中非技术能力的表征及多源定量评价研究”(71774010)
国际科研合作基金项目“基于属性异构网络表示学习的技术交易推荐方法研究”(2021B35)。
关键词
属性异构网络
网络表示学习
专利交易推荐
attribute heterogeneous network(AHN)
network representation learning
patent transaction recommendation
