Normalized auto-encoder based on biased walk for network representation learning.

Network representation learning based on graph auto-encoder has been the mainstream technique for learning embeddings in low-dimensional space. Graph auto-encoders learn the network representations by minimizing the loss between the generated and actual samples. However, limited information on the contextual structure of nodes is captured during the generation of samples. Furthermore, the dimensionality of the features is altered in the process of feature aggregation, resulting in impaired feature information in the network. In order to solve these two problems, normalized auto-encoder based on biased walk for network representation learning is proposed. This model can combine the advantages of shallow neural network and deep neural network to solve these two problems, and contains two key design components. One is the biased walk regularization module, a regularization term of shallow neural network, which captured broader and deeper context structure information of nodes. The other is the normalized regularization module of features, a regularization term of deep neural network, which smoothed the feature difference of nodes, and thus improved the feature information aggregation capability of convolution layer. This paper conducted experiments on three datasets Cora, CiteSeer and PubMed and compared them with other state-of-the-art network representation learning models. The results show that the proposed model can improve the link prediction results by 0.7% on AUC and 0.9% on AP to the state-of-the-art algorithms. In addition, the ablation experiments are conducted for two key design components, which can demonstrate the necessity of the proposed two regularization terms. [Display omitted] • A novel network representation learning model on graph auto-encoder is proposed. • The biased walk regularization module learns a wider range of contextual structural information. • The normalized regularization module improves the feature information aggregation capability. [ABSTRACT FROM AUTHOR]