摘要： 在多数属性网络嵌入算法中，拓扑结构的设计只考虑节点间直接链接，而未考虑节点间间接链接及不同节点的共同链接比，导致不能充分提取网络真实拓扑特征。针对该问题，提出一种基于稀疏自编码器的属性网络嵌入算法SAANE。根据网络拓扑提取二级邻居和共同邻居比并将其融入节点文本属性信息，对融合后的向量通过训练最优稀疏自编码网络得到节点低维嵌入向量。在5个真实网络上进行聚类和分类，实验结果表明，与DeepWalk、Node2Ves、LINE等8种主流算法相比，SAANE的聚类结果最优，NMI值平均提高5.83%，分类准确率平均提高4.53%。

关键词: 网络嵌入向量, 网络表示学习, 稀疏自编码器, 属性网络, 复杂网络

Abstract: Most of attribute network embedding algorithms only consider the direct links between nodes when designing topology structure,not the indirect links or the common link ratio of different nodes,which leads to the inadequate extraction of the real network topology characteristics.To solve this problem,an attribute network embedding algorithm based on sparse auto-encoder,SAANE,is proposed.The second-level neighbor-to-common neighbor ratio is extracted according to the network topology.On this basis,the text attribute information of the node is fused,and the fused vector is trained to obtain the low-dimensional embedding vectors of the node by training the optimal sparse self-coding network.Results of clustering and classification experiments on five real networks show that,SAANE outperforms DeepWalk,Node2Ves,LINE and other five mainstream algorithms in terms of clustering performance,increasing the average NMI value by 5.83% and the average classification accuracy by 4.53%.

Key words: network embedding vector, network representation learning, sparse auto-encoder, attribute network, complex network

中图分类号: 

• TP18