摘要： 现有网络嵌入算法大多只保留网络的微观结构信息，忽略了网络中普遍存在的社区结构信息。为提高网络表示质量，提出一种保留社区结构信息的网络嵌入算法PCNE。通过最大化节点之间的一阶和二阶相似性，对网络的微观结构进行建模，同时通过分解可反映网络社区结构信息的社区结构嵌入矩阵，对网络的社区结构信息进行建模。将构建的2个模型融合到统一的联合非负矩阵分解框架中，结合相似度矩阵和社区隶属度矩阵得到融合社区结构信息的节点表示向量。在5个真实公开数据集上进行节点分类实验，结果表明，与DeepWalk、Node2vec、LINE算法相比，PCNE可使Micro-F1值提升0.96%~13.1%，验证了算法的有效性。

关键词: 网络嵌入, 社区结构, 非负矩阵分解, 网络表示学习, 复杂网络

Abstract: Most existing network embedding algorithms only retain the micro-structure information of the network, but ignore the community structure information which is important in networks.In order to incorporate the community structure information into the network embedding to improve the quality of network representation, a network embedding algorithm preserving community information in network embedding, named PCNE that preserves community structure information is proposed.The micro-structure of the network is modeled by maximizing the first-order and second-order similarity between nodes, and then the community structure information of the network is modeled by factorizing the community structure embedding matrix which can reflect the community structure information of the network.Under the joint supervision of the similarity matrix of the micro-structure and the community membership matrix of the meso-structure, PCNE obtains the node representation vectors that fused the community structure information by merging the both into a unified joint non-negative matrix factorization framework.The performance of the PCNE algorithm is evaluated by node classification experiments on five real public datasets and compared with other five state-of-the-art embedding models.Experimental results show that the proposed method improves the Micro-F1 by 0.96%~13.1% compared with classical algorithms such as DeepWalk, Node2Vec and LINE, thus verifying the effectiveness of PCNE.

Key words: network embedding, community structure, non-negative matrix factorization, network representation learning, complex network

中图分类号: 

• TP18