基于异构信息网络嵌入的专利技术主体间交易推荐模型

考虑专利技术主体间技术邻近、地理邻近、共申请关系、引证关系、经济圈效应、主体类型邻近、主体间从属关系这7种因素对交易的影响,构建由4类节点、10类关系组成的异构信息网络,设计基于元路径与元结构的异构关系遍历算法获取主体间关系序列。以关系序列为语料,构建基于网络嵌入的异构信息网络主体间交易推荐模型(PSR-vec),采用基于Huffman树的Skip-Gram方法进行网络嵌入训练,计算主体向量间相似度以实现交易推荐。通过2012-2018年电子信息领域专利数据的实证研究得出:第一,PSR-vec模型相比DeepWalk、node2vec与PathSim等方法,推荐精度大幅提高,达到82.4%;第二,融合多个元路径与元结构特征的推荐与单一特征相比,推荐精度大幅提高;第三,基于ρ2以及改进的元结构S4、S6、S8、S10、S12、S14的推荐结果均高于基于ρ1以及改进的元结构S3、S5、S7、S9、S11、S13的推荐精度,说明基于主体间转让技术邻近性的推荐精度更高;第四,在技术邻近元路径基础上分别融合主体间共申请、引证、从属、经济圈效应这4类邻近关系获得元结构并进行推荐,推荐精度均显著提高,而融合地理邻近、类型邻近2类关系后推荐精度有所降低,说明地理邻近、类型邻近对交易的促进作用不明显;第五,基于PSR-vec模型的推荐结果包括具有控股和供应等紧密关系的主体,也包括关系松散的主体,推荐结果具有多样性。本研究为专利技术主体间的有效对接提供了决策方法。

Heterogeneous Information Network Embedding for Patent Technology Supply-Demand Trade Recommendations among Subjects

We constructed a heterogeneous information network(HIN) consisting of four types of nodes and ten types of relationships after considering the impact of seven parameters(technological proximity, geographical proximity, co-application relationship, citation relation, economic circle effect, subject type proximity, and subordinate relation among subjects) on trade. We then proposed heterogeneous relationship traversal algorithms based on the meta path and meta structure and obtained subject-relationship sequences based on multi-relational mapping. Based on the relation sequence corpus, we constructed a model of patent technology subject trade recommendations using network embedding(PSR-vec). We then trained the model using the Skip-Gram method of the Huffman tree to obtain a subject vector space representation. Finally, we calculated the similarity between subjects vectors to formulate a trade recommendation. Through empirical research on patent data in the electronic information field from 2012 to 2018, we found that, first, with an accuracy rate of82.4%, the PSR-vec model had greater accuracy compared to the DeepWalk, node2 vec, and PathSim methods. Second, the recommendation accuracy of the combination of multiple meta paths and meta structures was higher than that of the single meta path or meta structure recommendation. Third, the accuracy of the recommendation results based on ρ2 and the meta structures S4, S6, S8, S10, S12, and S14 is higher than the accuracy of those based on ρ1 and improved meta-structures S3,S5, S7, S9, S11, and S13. This shows that the recommendation accuracy based on technological proximity among subjects is higher. Fourth, the recommendation accuracy is significantly improved for the meta structures with co-application, citation, subordinate relation, and economic circle effect on the basis of technological proximity, while the recommendation accuracy is lower when combining geographical proximity and subject type proximity. This indicates that geographical proximity and subject type proximity do not promote patent technology trade. Fifth, the recommendation subjects included the controlling relationship, the upstream and downstream relationship of the supply chain, and loosely related subjects, which reflected the validity and novelty of the recommendation results. This research provides a decision-making method for effective docking between subjects.