基于深度学习的大规模图数据挖掘

发布时间：2018-09-11 05:56

【摘要】：随着大数据思维逐渐深入人心以及深度学习的广泛研究和应用,图结构逐渐被用来表征现实世界中大规模的、错综复杂的数据,而深层挖掘大规模图数据内部隐含的信息也逐渐成为了研究的热点。在信息爆炸的时代,传统的基于关键字匹配的搜索引擎已经难以满足用户希望迅速、准确、简便地获取信息的需求,为此知识图谱通过建立基于语义的信息实体图来满足人们新的查询需求。本文首先通过回顾学者、科研机构及公司对知识图谱的研究内容,对知识图谱的发展和构建方法作了全面的介绍,包括知识图谱概念的起源、发展以及最终形成过程;构建知识图谱的数据来源;构建过程中涉及的方法,包括本体和实体的抽取,图谱的构建、更新、维护,以及面向知识图谱的内部结构挖掘和外部扩展应用。最后,对知识图谱的未来发展方向和面临的挑战作了展望。针对大规模图数据挖掘面临的计算复杂、数据稀疏的问题,本文在word2vec算法基础上进行改进设计了一种基于深度学习的网络表示学习算法,通过将图结点表示为低维向量为图数据挖掘工作中能够使用成熟的机器学习算法和线性代数的理论和工具提供了可能。该算法针对图结点的多标签分类任务,利用部分标签信息指导在结点间游走的过程,然后使用逻辑回归分类模型对结点的特征表示进行多标签分类。实验结果显示通过有指导地游走,标签分类准确率有明显提升。另外,本文利用网络表示学习算法得到的图结点的向量表示设计了一种生成边特征表示的组合方法,同时通过构建深度置信网络的分类模型,实现了对复杂网络的链路预测。
[Abstract]:With the extensive research and application of big data's thinking and deep learning, the graph structure is gradually used to represent the large-scale and complicated data in the real world. And deep mining the hidden information inside the large scale map data has gradually become the hot spot of research. In the era of information explosion, the traditional search engine based on keyword matching has been difficult to meet the needs of users who want to obtain information quickly, accurately and easily. Therefore, the knowledge map can meet the new query needs by building semantic information entity graph. Firstly, by reviewing the research contents of knowledge atlas by scholars, scientific research institutions and companies, this paper gives a comprehensive introduction to the development and construction methods of knowledge atlas, including the origin, development and final forming process of the concept of knowledge atlas; The methods involved in constructing knowledge map include ontology and entity extraction, graph construction, updating, maintenance, and knowledge map oriented internal structure mining and external extension application. Finally, the future development direction and challenges of knowledge map are prospected. Aiming at the problem of complex computation and sparse data in large-scale graph data mining, a network representation learning algorithm based on deep learning is proposed in this paper, which is improved on the basis of word2vec algorithm. By representing graph nodes as low-dimensional vectors, it is possible to use mature machine learning algorithms and linear algebra theories and tools in graph data mining. According to the multi-label classification task of graph nodes, the algorithm uses partial label information to guide the process of walking between nodes, and then uses the logical regression classification model to classify the feature representation of nodes. The experimental results show that the accuracy of label classification is significantly improved by guided walking. In addition, using the vector representation of graph nodes obtained by network representation learning algorithm, a combination method of generating edge feature representation is designed. At the same time, the link prediction of complex networks is realized by constructing a classification model of depth confidence networks.
【学位授予单位】：南京邮电大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP311.13;TP181

【参考文献】