Graph Convolutional Extreme Learning Machine Autoencoder for Graph Embedding

Abstract

The purpose of graph embedding is to encode the known node features and topological information of graph into low-dimensional embeddings for further downstream learning tasks. Graph autoencoders can aggregate graph topology and node features, but it is highly dependent on the gradient descent optimizer with a large iterative learning time, and susceptible to local optimal solutions. Thus, we propose Graph Convolutional Extreme Learning Machine Autoencoder. To address the limitation that the extreme learning machine autoencoder cannot use topological information, the graph convolution operation is introduced between the input layer and the hidden layer to improve the representation ability of the graph embedding obtained. Experiments of link prediction and node classification on 5 real datasets show that our method is effective.