基于平滑l1范数的深度稀疏自动编码器社区识别算法

Sparse autoencoder community recognition algorithm based on smoothed l1 norm

大数据时代,利用传统的社区发现算法对大规模复杂网络进行社区结构挖掘显得愈发困难,准确率也较低。因此,提出一种基于平滑l1范数的深度稀疏自编码器社区发现算法(l1-ECDA)。该算法首先采用基于s跳的方法对网络图的邻接矩阵进行预处理;然后构建基于平滑l 1范数的深度稀疏自编码器,并通过训练网络图相似度矩阵得到低维特征矩阵;最后采用K-means算法对低维特征矩阵进行聚类得到网络社区结构。通过在仿真网络与真实网络数据集上的实验表明,l1-ECDA有效提高了社区识别的准确率,且准确率比DBCS算法平均高4%,比DeepWalk和CoDDA算法平均高5.4%。

In the age of big data,it is increasingly difficult to make the community structure mining of large-scale complex networks by using the traditional community discovery algorithm and the accuracy rate is low.Therefore,this research came up with l1-ECDA,a community discovery algorithm for deep sparse self-encoder based on smooth l1 norm.This algorithm preprocessed the adjacency matrix of the network diagram with the method based on s jump.Then it established the deep sparse self-encoder based on smooth l1 norm and got the low dimensional characteristic matrix by training the similarity matrix of the network graph.Finally,it got the network community structure by clustering the low-dimensional feature matrix through the K-means algorithm.Experiments on simulated network and real network data set show that l1-ECDA improves the accuracy of community recognition effectively.Its accuracy rate is 4% higher than the DBCS algorithm on average,and is 5.4%higher than DeepWalk algorithm and CoDDA algorithm on average.