Network Embedding Method Based On Extractive Summary

Abstract

Redundant or low quality sampling sequences are used in existing network embedding methods based on random walk. A network embedding method based on extractive summary is proposed to generate high-quality node embedding. A selective gate network is used by the role of the node in the overall sequence. A decoder based on extractive abstract is designed by prediction and sampled condition of the node. Firstly, by using the control characteristics of the selective gate network, the hidden state vectors containing the attribute information are filtered. The environment vectors that can effectively represent the key information of nodes are acquired. It achieves the extraction of important information of the node. Furthermore, the environment vector is decoded by the extractive-abstract-based decoder. The redundant nodes in the original sampling sequence are removed, which further improves the classification accuracy. With the datasets of Cora, Citeseer and Wiki, the proposed method is applied to network node classification, and outperforms several mainstream baseline methods.