Network Representation Learning Algorithm Combined with Node Text Information

Abstract

In a network data analysis task, network representation learning play an important role. It can obtain a low-dimensional representative vector from high dimensional features of nodes which will promote the efficiency of downstream analysis task. Many algorithms mainly based on the structure information of the network but for a complex network, there is a lot kinds of feature information available for representation learning which has an important impact on network data analysis task. Therefore, we propose a network representation learning algorithm that combines node text information. Before our algorithm, we encode the information of the nodes including structure and text, and the strategy of fusion is adopted in advance which splice structure and text information. And then we feed the splicing results into the neural network which is a multi-layer feedforward neural network with a tower structure to learn the nonlinear structure information of complex networks. Finally, we design the structure loss function that preserve structure similarity and text loss function that preserve text information respectively. And we use the joint training to make two aspects information interact continuously. We conduct experiment to evaluate our algorithm. The results on two real word networks show our algorithm has better performance compared with traditional algorithms.