基于图的社交网络匿名群体识别关键技术

The illegal action of special organizations which hidden in the large scale social network threaten the cyberspace and national security seriously. It is significance for detecting and recognizing such anonymous organization. Sub-graph classification is an effective method of anonymous organization recognition. But the high dimensions graph data training leads to classification algorithm low efficiency. Based on the network structure of sub-graph, this project tries to analysis the structure features of sub graph, construct the classification model. And the research is focus on sub-graph formalization, sub-graph classification feature construction, and graph classification algorithm. This research will solve the recognition of anonymous organization in social network. Specifically, the main research topic content includes: 1) Research on social network groups’ division method, identify independence group structure from large social networks. 2) Research on social network group similarity comparison method; find out the similar organization structure in large-scale network. 3) Research on feature selection method of social network community recognition, construct feature vector for classification. 4) Research on group classification model of social networking and parallelization techniques, realizing the anonymous group recognition based on graph classification, and prompting the warning efficiently..While the research is completed, it will perfect the graph classification methods based on the network structure, identify the special organization which hidden in large scale social network in real time, and will support the Counterterrorism and Cybersecurity. And it also promotes the healthy development. While the research is completed, it will perfect the graph classification methods based on the network structure, identify the special organization which hidden in large scale social network in real time, and will support the Counterterrorism and Cybersecurity.

隐匿在社交网络中的个别组织的非法行为严重危害着网络空间安全和国家安全，发现和识别此类匿名组织具有现实意义。子图分类是实现匿名组织识别的有效手段，但高维的图数据学习使得分类算法效率低下。本课题拟以社交网络中子图结构信息为基础，通过构建基于图的分类模型，从子图形成、子图分类特征构建、 子图分类三个层面进行研究，解决社交网络匿名组织识别问题。具体而言：1)研究社交网络中匿名群体发现方法，实现从大规模社交网络找出独立群体和交叉群体；2)研究社交网络中群体相似性比较方法，找出大规模网络结构中频繁出现的类似组织；3)研究社交网络中群体识别的特征选择方法，构建特定组织的结构信息特征向量；4)研究社交网络的群体分类模型及并行化技术，实现基于图分类的匿名组织识别，提升预警时效。通过本课题的研究，完善基于网络结构信息的图分类方法，实时发现隐藏在广袤的社交网络中的个别组织，为网络反恐提供技术支持，保障国家安全。

隐藏在浩瀚的社交网络大数据中的恐怖分子往往是匿名存在，挖掘此类组织的基础只有网络拓扑关系。以社交网络中图结构信息为基础，通过构建基于图的分类模型，从而发现隐匿组织具有重要意义。本项目主要开展的研究有：采用并行的标签传播算法来实现社交网络中的群体发现，使得群体发现具有较高的可扩展性和准确性；采用基于局部信息和全局信息的节点影响力计算方法来进行重叠社区划分，实验表明该算法发现的网络社区结构更加精确；采用基于节点影响力的子图匹配算法来实现频繁子图挖掘，采用基于查询图分解的子图匹配改进算法解决查询图较大时子图匹配效率较低的问题，并利用Spark框架进行并行化处理，提升子图匹配效率；考虑频繁子图与非频繁子图对图类别的判别力，采用去除低判别性子图和部分不频繁子图来重构原始图集，组建由低判别力频繁子图集合、高判别力频繁子图集合、低判别力非频繁子图资格和高判别力非频繁子图集合组成的子图分类特征；采用图结构嵌入方式构建反映图和图与图之间的相对关系的图向量，使用支持向量机分类器对图进行分类取得较好的效果；采用基于重要节点结构特征和空间卷积的图分类方法，从局部和全局两个方面选择节点特征，保留了更多的图结构信息，根据邻域图的分类来确定整个图的分类结果，优于一些经典图核方法和基于深度学习的空间卷积模型方法。依托本项目，已经发表文章22篇，其中SCI文章21篇，申请密切相关的专利9项，培养博士7人（其中毕业2人）、硕士16人（其中毕业9人）。依托于本项目的部分研究成果，申请了科技部国家重点研发计划政府间国际科技创新合作重点专项。