【作者】 刘鑫；

【导师】 刘维东；

【作者基本信息】 内蒙古大学 ， 计算机科学与技术， 2021， 硕士

【摘要】 随着全球知识产权竞争的加剧,我国在经济发展“新常态”下大力实施创新驱动发展战略,国民科技创新能力大幅提升,知识产权意识显著提高。专利作为知识产权中的一种重要形式,它潜在的经济价值一直备受瞩目。专利估值作为实现专利技术向专利产业化发展的重要环节,对于专利交易、专利投资、专利转化等而言是极其重要的。然而,目前自动化专利估值尚处于初步发展阶段,以下挑战性问题有待解决,具体包括:如何表示估值对象,如何构建估值场景,以及如何生成专利价值。为解决上述挑战性问题。首先,为了表示估值对象,本文侧重于专利的结构化部分和非结构化的文本内容,并从中提取一些特征来表示一个估值对象。然后,由于专利数据与市场数据存在复杂的知识关联,专利的价值隐含在这些复杂的知识关联当中。因此,本文从知识关联的角度出发,构建了一个由专利数据与市场数据组成的异构知识关联网络,将其作为估值场景,为本文的研究提供了一个重要的计算基础。在该估值场景中,通过专利与其关联的市场数据之间的相互影响,使得专利的价值随其关联的市场数据的变化而波动,产生价值联动效应。最后,在表示好估值对象和构建好估值场景的基础上,本文从有监督学习和无监督学习两个角度来对专利价值的生成过程进行研究。1)有监督学习:本文提出了基于贝叶斯图卷积神经网络的专利估值模型。在该模型中,本文使用贝叶斯方法,将可观测的估值场景视为一个来自随机图参数族的实现。然后,本文针对随机图参数的后验进行推理,进而对其进行社区划分,生成新的估值场景。最后在生成的估值场景上使用贝叶斯图卷积神经网络生成专利价值。2)无监督学习:本文提出了基于概率图的专利估值模型。在该模型中,根据专利和市场数据的自身特征形成它们价值的先验分布。然后,将它们置身于估值场景中,提出价值链接假设,通过概率生成过程,构建概率图模型。最后,使用变分推理算法近似推理专利价值的后验分布。在专利数据集上,本文将提出的两种模型与最新模型进行对比,实验结果表明本文提出的模型优于对比模型。更多还原

【Abstract】 With the intensification of global intellectual property competition,China has vigorously implemented the innovation-driven development strategy under the "new normal" of economic development,so the national scientific and technological innovation capabilities have been greatly improved,and the awareness of intellectual property rights has increased significantly.As an important form of intellectual property rights,patents have always attracted attention to their latent economic value.As an important bridge between the development of patent technology and patent industrialization,patent valuation is extremely important for patent transactions,patent investment,and patent transformation.However,the current automation patent valuation is still in the preliminary stage.The following challenging issues need to be resolved: how to represent the valuation object,how to construct a valuation scenario,and how to generate patent value.To solve those challenges,it is necessary to follow these steps.First,to represent the valuation object,this article focuses on the structured part and unstructured text content of the patent,and extracts some features from it to represent a valuation object.Then,because patent data and market data have complex knowledge associations,the patent value is implicit in these complex knowledge associations.Therefore,from the perspective of knowledge relevance,this study constructs a heterogeneous knowledge association network composed of patent data and market data,and uses it as a valuation scenario to provide an important calculation basis for the study of this study.In this valuation scenario,through the mutual influence between the patent and its associated market data,the value of the patent fluctuates with changes in its associated market data,resulting in a value linkage effect.Finally,on the basis of expressing good valuation objects and constructing good valuation scenarios,this article studies the generation process of patent value from two perspectives: supervised learning and unsupervised learning.1)Supervised learning: We propose a patent valuation model based on Bayesian graph convolutional neural network.In this model,we use Bayesian methods to treat the observable valuation scenario as an implementation from a random graph parameter family.Then,we reason about the posterior of the random graph parameters,and then divide them into communities to generate new valuation scenarios.Finally,the Bayesian graph convolutional neural network is used to generate patent value on the generated valuation scenario.2)Unsupervised learning: We propose a patent valuation model based on probability graphs.In this model,the prior distribution of their value is formed based on the characteristics of patents and market data.Then,put them in the valuation scenario,propose the value link hypothesis,and build the probabilistic graph model through the probabilistic generation process.Finally,variational inference algorithm is used to approximate the posterior distribution of patent value.In terms of the patent dataset,two models proposed in this thesis have been confirmed via comparing with the state-of-the-art model.In the evaluation and measurement,the models in this thesis outperform comparised models.更多还原