基于图核的复杂数据异常检测算法研究.docVIP

摘要随着信息技术、互联网、云计算等产业的飞速发展，数据的采集、存储与传输变得更加便捷，使得人们迫切的需要从大量的数据中挖掘出重要的信息，数据挖掘得以飞速发展，异常检测则是数据挖掘中的重要组成部分。在海量的复杂数据中存在着一类图数据，他们呈现出一定的拓扑、几何特性，如化学分子、社交网络等。本文的研究内容就是图数据的异常检测，并根据不同数据集提出了相关算法。 首先，本文利用图模型、邻接矩阵等方法对图数据进行描述，研究了图数据后选用了图核方法作为图数据之间的相似性度量，并利用KPCA方法对图核处理后的数据进行降维处理。 其次，针对训练过程中使用的数据集相对不平衡状态，对图核处理并降维后的数据集采用对不同类别引进不同惩罚因子的加权SVM、加权SVDD-Negative方法，提出了基于图核的相对不平衡数据集异常检测算法，并在小数据集上进行了算法的分析，实验结果表明本文提出的算法在数据量相对不平衡的状态下具有一定的异常检测效果。 接着，针对绝对不平衡状态即单类问题，对图核处理并降维后的某一类别数据利用单类分类器方法如One Class SVM、单类SVDD方法进行模型的训练，随后对未知属性的数据进行异常检测，提出了基于图核的绝对不平衡数据集异常检测算法，实验结果表明本文提出的算法对绝对不平衡数据集具有一定的异常检测效果。 再次，利用在实际应用中的两个复杂的数据集对本文提出的算法进行了性能验证，表明本文提出的算法具有一定的实际意义。 关键词：图核、图数据、数据不平衡、异常检测   ABSTRACTWith the rapid development of information technology, the rapid progress of internet, cloud computing and other emerging industry, data collection, storage and transmission have become more convenient. People urgent to dig out the useful information from large amounts of data, data mining was developed and anomaly detection is an important part of data mining. There is a kind of graph data in the complex data, and they have the topological?characteristics, such as chemical, social networks, etc. In this paper we make a research on anomaly detection of graph data, and put forward the related algorithm according to different datasets. Firstly, we use the method of graph model or adjacency matrix to describe this graph data, and use the graph kernel method as the similarity measure between the graph data, and using KPCA method to reduce the dimension of the data processing by graph kernel. Secondly, we use the different penalty factor according to different categories of data to solve the relatively imbalances datasets, such as SVM-Weights, SVDD-Negative Weights, and put forward the relatively imbalances anomaly detection algorithm based on graph kernel. We test our algorithm on small data set and the results show that the proposed algorithm has a certain abnormal detection effect on the relative unbalance data set.