Abstract:

Attributed network representation learning aims to combine structure information and attribute information to learn a uniform vector representation for the nodes in the network. The existing method cannot enhance its complementary structural information enough when learning attribute information, thus affecting the final representation. In order to improve the representation quality, this paper proposes a structure-enhanced attributed network representation learning method. More specifically, this method extracts attribute information that enhances the global adjacency of the network through an autoencoder based on the normalized adjacency matrix and attribute matrix, the skip-gram model is used to capture local structure information, and a joint loss function is introduced to make the structure information and attribute information represent in the same vector space. Extensive experiments of node classification and link prediction on three real attribute networks demonstrate a superior performance of the proposed method over state-of-the-art for network representation learning.

Key words: network representation learning, attributed network, deep learning

摘要：

属性网络表示学习旨在结合结构信息与属性信息为网络中的节点学习统一的向量表示。现有的属性网络表示学习方法在学习属性信息时与其互补的结构信息增强不足，从而影响最终表示。针对这一问题，提出一种结构增强的属性网络表示学习方法，以提高表示质量。该方法基于网络归一化邻接矩阵和属性矩阵通过自动编码器提取增强网络全局结构特性的属性信息，使用skip-gram模型捕捉局部结构信息，引入一个联合损失函数使结构信息与属性信息在同一向量空间中得以表示。在三个真实属性网络数据上进行节点分类和链路预测实验，效果较目前流行的网络表示学习方法优势明显。

关键词: 网络表示学习, 属性网络, 深度学习