并行时序网络预测模型研究及其在异常社团检测中的应用

Abnormal community detection is an important research task in dynamic community discovery. It has important theoretical significance and practical value in revealing the evolution mechanism of online social networks and mastering the trend of network development. However, the lack of credible samples caused by the sparsity of network information and the high computational complexity caused by the large scale network data seriously restrict the performance improvement of abnormal group detection in online social networks. In view of the above bottleneck problems, the project proposes a multi-level features extraction algorithm to achieve the purpose of indirectly increasing the network low-dimensional trusted samples by combing the traditional feature extraction algorithm and the theory of network embedding. In addition, in view of the advantages of cyclic neural network, a parameters shared parallel prediction learning model of sequential network is established to predict the evolution process of the communities, then realizing the rapid identification of abnormal groups according to the network evolution. The specific contents are as follows: 1) propose a community structure features extraction algorithm which based on network embeddedness for dynamic network data; 2) establish a parallel prediction learning model of sequential network to predict the evolution process of the communities adaptively; 3) propose a similarity measure which based on the evolution states and the evolution processes of the communities to realize the anomaly groups detection in WeChat network. It also provides new research ideas and more effective solutions for the task of meso-structure prediction in the online social networks.

异常群体检测是动态社区发现的重要研究内容，对于揭示在线社会网络的演化机理，掌握网络发展态势具有重要理论意义及实用价值。然而网络信息稀疏性造成的可信样本匮乏和大规模数据处理计算复杂性高的现实情况，严重制约了在线社交网络异常群体检测性能的提高。针对以上瓶颈问题，本课题将传统的特征提取算法和网络嵌入思想相结合，提出基于网络嵌入的多层次特征提取算法，达到间接增加网络低维可信样本的目的；此外，鉴于循环神经网络对时序数据处理的优势，建立参数共享的并行时序网络预测模型，预测网络社团演化情况，实现基于网络演化的异常群体快速识别。具体内容包括：1）针对动态网络数据提出基于网络嵌入的社团特征提取算法；2）建立并行时序网络预测模型，自适应的预测社团演化过程；3）根据社团演化状态和演化过程提出相似性度量准则，实现微信网络的异常团体检测。为在线社交网络的中观结构预测任务提供新的研究思路和更有效的解决办法。

异常群体检测是动态社区发现的重要研究内容，对于揭示在线社会网络的演化机理，掌握网络发展态势具有重要理论意义及实用价值。然而网络信息稀疏性造成的可信样本匮乏和大规模数据处理计算复杂性高的现实情况，严重制约了在线社交网络异常群体检测性能的提高。针对以上瓶颈问题，本课题将传统的特征提取算法和网络嵌入思想相结合，提出基于网络嵌入的多层次特征提取算法，达到间接增加网络低维可信样本的目的；此外，鉴于循环神经网络对时序数据处理的优势，建立参数共享的并行时序网络预测模型，预测网络社团演化情况，实现基于网络演化的异常群体快速识别。具体内容包括：1）针对动态网络数据提出基于网络嵌入的社团特征提取算法；2）建立并行时序网络预测模型，自适应的预测社团演化过程；3）根据社团演化状态和演化过程提出相似性度量准则，实现微信网络的异常团体检测。为在线社交网络的中观结构预测任务提供新的研究思路和更有效的解决办法。研究成果在本领域重要期刊和会议上发表论文 4 篇，其中 SCI 1 区论文 2 篇；申请国家发明专利 3 项；联合培养硕士生4名，其中毕业 2名。