基于图流在线非负矩阵分解的社团检测

Graph Streams Community Detection via Online Nonnegative Matrix Factorizations

针对现有的在线社团检测方法大多仅从增量相关的节点和边出发,难以有效挖掘社团结构的动态变化特性问题,提出了一种基于图流在线非负矩阵分解的社团检测方法.首先将网络中持续到达的图数据按照流式数据进行存储和预处理,然后借鉴梯度下降思想,采用在线非负矩阵分解架构,根据不同时刻达到的图流序列,实时迭代更新社团归属矩阵,并通过有效的学习率和缓存策略设置,保证了图流处理的收敛性和合理性.实验结果表明,相比于已有在线社团检测方法,该方法具备更高的社团检测精度.

While the existing online community detection methods mostly only deal with the nodes and edges from the increment part,which are difficult to effectively detect the dynamic changes in the community structure. Based on this,a newmethod for the detection of flowgraphs based on online non negative matrix factorization（ ONMF） is proposed. Firstly,our method put graph data into the cache as continuous streams to deal with. Then,our method iteratively updates the existing community belonging matrix real-time using online nonnegative matrix decomposition architecture and by means of the projected gradient descent theory. Lastly,through effective learning rate and cache strategy setting,our method ensures the convergence and rationality of graph stream processing. Experiments on real network data sets showthat ONMhas a higher community detection quality compared with the existing methods.