大规模多源异构社交网络数据驱动的异常群体识别

Multi-source heterogeneous social network data in cyberspace has the challenges of large-scale, high-dimensional heterogeneity, real-time dynamics, and low number of labeled samples, which makes it difficult to recognize the terror-related and rumor-related abnormal groups. This project takes large-scale multi-source heterogeneous social network data as the research object, and differentiates the abnormal group recognition problem into the abnormal group discovery problem in the heterogeneous social network and the abnormal group recognition problem based on its attribute features and topology features. We propose a three-stage solution of "group discovery – group clustering – group recognition". We will target the key issues of the parallelization of abnormal group discovery algorithms in large-scale heterogeneous social networks, abnormal group clustering based on high dimensional heterogeneous information fusion, and abnormal group recognition and interpretation based on multi-machine learning methods. This project aims to form a set of methods and theories for large-scale multi-source heterogeneous social network data-driven abnormal group recognition. This research would solve the hidden and adversarial problems of abnormal groups, and provide the theoretical basis and technical support for the violent and terrorist-related abnormal group recognition in cyberspace, which is of great significance for safeguarding the national cyberspace security.

网络空间中大规模多源异构社交网络数据具有的大规模性、高维异构性、实时动态性以及标记样本少等特性导致对涉暴涉恐、谣言传播等异常群体很难进行识别。本项目以大规模多源异构社交网络数据为研究对象，将异常群体识别难题分化为异构网络中异常群体发现问题以及基于其属性特征与拓扑特征的异常群体识别问题，提出“群体发现—群体聚类—群体识别”三阶段解决思路；据此解决大规模多源异构社交网络中异常群体发现算法的并行化、基于高维度异构信息融合的异常群体聚类以及基于多类机器学习方法有机结合的异常群体识别与解释等关键问题。本项目旨在形成一套完善的大规模多源异构社交网络数据驱动的异常群体识别的方法与理论，解决异常群体的隐蔽性与对抗性难题，为网络空间内涉暴涉恐异常群体的识别提供理论依据和技术支持，对于保障国家网络空间安全具有重要意义。

本项目在异常群体结构发现和挖掘的基础上，结合群体的属性特征及拓扑特征，利用多类机器学习方法的有机结合对异常群体进行有效识别和解释。项目研究内容层层递进，以实现大规模多源异构社交网络数据驱动的异常群体发现、聚类及识别，研制出基于多源异构数据的异常人物画像及定位原型系统，并在国家某安全部门、人民网、西安市公安局技术侦察支队等国家重要安全部门开展了实证测试与验证，处理了大量用户的多源异构数据，取得了属性关联补全率超过70%、群体发现准确率达到80%、关注目标锁定准确率高于80%，召回率高于70%等实战效果，大规模多源异构社交网络数据挖掘及异常群体识别等问题提供重要的理论依据和技术支持，从而大大提高网络空间安全的保障水平。.项目执行期间，基于本项目研究内容发表相关论文7篇，其中国际期刊论文5篇、CCF A类国际会议论文1篇、中文CCF A类期刊1篇，授权或申请发明专利和软件著作权4项。课题负责人入选2020年西安交通大学“青年优秀人才计划”A类，获得2021年陕西省优秀博士论文、传播内容认知国家重点实验室原创性成果奖励（一等）等荣誉；指导学生获得1名入选犀牛鸟精英人才计划、1名获得西安交通大学优秀硕士学位论文、1名获得“华为杯”第十八届中国研究生数学建模竞赛二等奖；