基于深度稀疏自动编码器的社区发现算法
发布时间：2018-10-15 18:24
【摘要】：社区结构是复杂网络的重要特征之一,社区发现对研究网络结构有重要的应用价值.k-均值等经典聚类算法是解决社区发现问题的一类基本方法.然而,在处理网络的高维矩阵时,使用这些经典聚类方法得到的社区往往不够准确.提出一种基于深度稀疏自动编码器的社区发现算法CoDDA(a community detection algorithm based on deep sparse autoencoder),尝试提高使用这些经典方法处理高维邻接矩阵进行社区发现的准确性.首先,提出基于跳数的处理方法,对稀疏的邻接矩阵进行优化处理,得到的相似度矩阵不仅能够反映网络拓扑结构中相连节点间的相似关系,同时还反映了不相连节点间的相似关系.然后,基于无监督深度学习方法构建深度稀疏自动编码器,对相似度矩阵进行特征提取,得到低维的特征矩阵.与邻接矩阵相比,特征矩阵对网络拓扑结构有更强的特征表达能力.最后,使用k-均值算法对低维特征矩阵聚类得到社区结构.实验结果显示:与6种典型的社区发现算法相比,CoDDA算法能够发现更准确的社区结构.同时,参数实验结果显示,CoDDA算法发现的社区结构比直接使用高维邻接矩阵的基本k-均值算法发现的社区结构更为准确.
[Abstract]:Community structure is one of the important characteristics of complex networks. Community discovery has important application value in studying network structure. The classical clustering algorithm such as k- mean is a kind of basic method to solve community discovery problem. However, when dealing with the high dimensional matrix of the network, the communities obtained by these classical clustering methods are often inaccurate. A community discovery algorithm based on deep sparse automatic encoder (CoDDA (a community detection algorithm based on deep sparse autoencoder),) is proposed to improve the accuracy of community discovery by using these classical methods to deal with high-dimensional adjacency matrix. First of all, a method based on hops is proposed to optimize the sparse adjacent matrix. The similarity matrix can not only reflect the similarity relationship between connected nodes in the network topology. At the same time, it also reflects the similarity between disconnected nodes. Then, based on the unsupervised depth learning method, the depth sparse automatic encoder is constructed, and the feature extraction of similarity matrix is carried out, and the low-dimensional feature matrix is obtained. Compared with the adjacent matrix, the feature matrix has a stronger ability to express the network topology. At last, we use k- mean algorithm to cluster the low dimensional feature matrix to get the community structure. Experimental results show that CoDDA algorithm can find more accurate community structure than six typical community discovery algorithms. At the same time, the experimental results show that the community structure discovered by the CoDDA algorithm is more accurate than that by the basic kmean algorithm using the high-dimensional adjacency matrix directly.
【作者单位】： 清华大学软件学院;
【基金】：国家自然科学基金(61373023)~~
【分类号】：O157.5
[Abstract]:Community structure is one of the important characteristics of complex networks. Community discovery has important application value in studying network structure. The classical clustering algorithm such as k- mean is a kind of basic method to solve community discovery problem. However, when dealing with the high dimensional matrix of the network, the communities obtained by these classical clustering methods are often inaccurate. A community discovery algorithm based on deep sparse automatic encoder (CoDDA (a community detection algorithm based on deep sparse autoencoder),) is proposed to improve the accuracy of community discovery by using these classical methods to deal with high-dimensional adjacency matrix. First of all, a method based on hops is proposed to optimize the sparse adjacent matrix. The similarity matrix can not only reflect the similarity relationship between connected nodes in the network topology. At the same time, it also reflects the similarity between disconnected nodes. Then, based on the unsupervised depth learning method, the depth sparse automatic encoder is constructed, and the feature extraction of similarity matrix is carried out, and the low-dimensional feature matrix is obtained. Compared with the adjacent matrix, the feature matrix has a stronger ability to express the network topology. At last, we use k- mean algorithm to cluster the low dimensional feature matrix to get the community structure. Experimental results show that CoDDA algorithm can find more accurate community structure than six typical community discovery algorithms. At the same time, the experimental results show that the community structure discovered by the CoDDA algorithm is more accurate than that by the basic kmean algorithm using the high-dimensional adjacency matrix directly.
【作者单位】： 清华大学软件学院;
【基金】：国家自然科学基金(61373023)~~
【分类号】：O157.5
【参考文献】
相关期刊论文 前4条
1 周小平;梁循;张海燕;;基于R-C模型的微博用户社区发现[J];软件学报;2014年12期
2 黄发良;张师超;朱晓峰;;基于多目标优化的网络社区发现方法[J];软件学报;2013年09期
3 淦文燕;赫南;李德毅;王建民;;一种基于拓扑势的网络社区发现方法[J];软件学报;2009年08期
4 杨楠,弓丹志,李_，
本文编号：2273437
本文链接：https://www.wllwen.com/kejilunwen/yysx/2273437.html
