信息网络安全 ›› 2016, Vol. 16 ›› Issue (6): 35-40.doi: 10.3969/j.issn.1671-1122.2016.06.006

• • 上一篇    下一篇

一种改进的图分割算法在用户行为异常检测中的应用

杨连群1, 温晋英1, 刘树发2, 王峰3   

  1. 1.天津市滨海新区公安局,天津300450
    2.天津市公安局,天津 300000
    3. 91655部队,北京 100036
  • 收稿日期:2016-05-15 出版日期:2016-06-20 发布日期:2020-05-13
  • 作者简介:

    作者简介: 杨连群(1978—),男,天津,高级工程师,硕士,主要研究方向为计算机算法;温晋英(1980—),男,天津,工程师,主要研究方向为计算机算法;刘树发(1978—),男,天津,工程师,硕士,主要研究方向为软件工程;王峰(1977—),男,山东,博士,主要研究方向为计算机系统结构。

  • 基金资助:
    国家高技术研究发展计划(国家863计划) [2013AA01A214]

An Improved Graph Partitioning Algorithm for User Behavior Abnormal Detection

Lianqun YANG1, Jinying WEN1, Shufa LIU2, Feng WANG3   

  1. 1. Binhai New Area Public Security Bureau, Tianjin 300450, China
    2. Tianjin Public Security Bureau, Tianjin 300000, China
    3. Army 91655, Beijing 100036, China
  • Received:2016-05-15 Online:2016-06-20 Published:2020-05-13

摘要:

基于模拟随机流的马尔科夫分类算法(Markov Cluster Algorithm,MCL)是一种快速且可扩展的无监督图分割算法, 在用户行为异常检测中具有广泛的应用,但时间复杂度为 O(N3),不利于处理海量数据。为提高分割质量,同时减少计算时间,文章提出了一种改进的 MCL 模型。采用调整互信息(Adjusted Mutual Information,AMI)指标,对不同时间的图分割结果的相似度进行比较,判断是否有异常发生。实验表明,相较多层图分割算法(METIS),文章所提出的改进的 MCL 模型具有以下优点: 1)无需事先规定聚类的数目; 2)不易被数据中的拓扑噪声所影响;3)适合处理长尾分布的数据 ; 4)在计算时间一定的情况下,能获得质量较高的分割结果 。

关键词: 图分割, 马尔科夫分类算法, 异常检测, 多层图分割算法

Abstract:

The MCL algorithm is short for Markov Cluster Algorithm, a fast and scalable unsupervised partitioning algorithm for graphs. MCL has been widely applied to anomaly detection. However, it requires O(N3) time, Which is no good for a wide range of data processing. In order to improve the quality of clustering and save time consumption, an improved MCL algorithm is proposed. With AMI (Adjusted Mutual Information) index, the similarities of clusters of different periods are compared to determine whether the abnormal behavior occurs. Compared with multilevel k-way partitioning scheme (METIS) for graphs, experiments show that the proposed MCL algorithm has following strengths: 1) Number of clusters not specified ahead of time. 2) Robust against noise in graph data. 3) Suitable for clusters with long tail distribution. 4) Produce better clustering results in case of certain time consumption.

Key words: graph partitioning, Markov Cluster Algorithm, abnormal detection, METIS

中图分类号: