面向专利大数据的机器学习与推理研究

With the rapid increase in patent applications, both creative examination for each patent and content analysis for total patent face unprecedented challenges. How to tackle the uncertainty of the relationship among the massive, heterogeneous, multi-modal and sparse patent data becomes a challenging problem for patent examination, patent analysis, patent planning and patent promotion. In view of this, we will develop theories and methods of machine learning and reasoning for big patent data, when we are undertaking a task of State Patent Bureau: the analysis and processing of massive data of biological characteristics. Firstly, to solve the uncertainty in the relationship among massive data, this project will innovatively propose methods for patent dimension reduction, patent feature selection and patent feature integration based on manifold learning and fuzzy reasoning. Secondly, by regarding a patent as a graphic node in probabilistic graphical model, semantic super-network will be introduced and constructed; To achieve efficient knowledge learning and knowledge reasoning, two novel algorithm will be respectively designed to extract meta-knowledge and implement the jointly inference. Finally, an integrated platform for in-depth semantic analysis of big patent data will be developed and we will conduct experiments on the biological characteristics of State Patent Bureau to demonstrate the effectiveness of the proposed theories and methods. The aim of this project is to provide a theoretical guidance for solving some compelling and unsolved problems in patent examination and services, to meet the in-depth semantic requirements of patent for patent examiners and patent users.

随着专利申请量的快速增长，对每一件专利创新性审查和整个专利库内容分析遇到前所未有的挑战，如何处理海量、异构、稀疏的专利大数据之间潜在的关联不确定性，是专利审核、分析、规划和推广人员面临的重要难题。针对我们正在承担的国家专利局生物特征大数据分析与处理任务，研究专利大数据的机器学习与推理理论和方法。首先，针对专利大数据关联、不确定性，研究基于流形学习和模糊推理的专利特征降维、选择和融合方法；其次，将每一个专利作为概率图模型的节点，研究构建基于概率图模型的语义超网络方法，以及萃取专利元知识和建立高效联合推理算法，实现对专利大数据的语义关联学习与知识推理；最后，针对国家专利局生物特征大数据库，研发的专利大数据深度语义分析平台，验证本项目提出理论和方法的正确性、有效性。预期研究成果有助于解决专利审核与服务中的亟待解决的难题，满足专利审查人员和广大专利用户对专利深层语义分析的需求。

如何处理海量、异构、稀疏的专利大数据之间潜在的关联不确定性，是专利审核、分析、规划和推广人员面临的重要难题。通过三年的研究，本项目完成的研究工作和取得的成果如下：（1）运用拉普拉斯特征映射理论，提出解决专利特征降维、选择和融合的方法,建立可用于多特征处理的技术手段和理论框架；（2）运用神经网络和概率图模型理论，提出构建专利语义网络方法；（3）基于熵和Keeney-Raiffa算法等理论，提出萃取专利元知识和建立高效联合推理算法。研究成果在国内外权威期刊上共发表学术论文22篇，其中SCI索引5篇、EI索引3篇，培养研究生7人。项目负责人凭借本课题研究成果以优异成绩通过中共河北省委、河北省政府“首批青年拔尖人才支持”期满考核，且获得 “河北省青年拔尖人才”第二个支持周期的经费资助等荣誉和奖励。