Graph deconvolutional networks

Abstract

Graphs and networks are very common data structure for modelling complex systems that are composed of a number of nodes and topologies, such as social networks, citation networks, biological protein-protein interactions networks, etc. In recent years, machine learning has become an efficient technique to obtain representation of graph for downstream graph analysis tasks, including node classification, link prediction, and community detection. Different with traditional graph analytical models, the representation learning on graph tries to learn low dimensional embeddings by means of machine learning models that could be trained in supervised, unsupervised or semi-supervised manners. Compared with traditional approaches that directly use input node attributes, these embeddings are much more informative and helpful for graph analysis. There are a number of developed models in this respect, that are different in the ways of measuring similarity of vertexes in both original space and feature space. In order to learn more efficient node representation with better generalization property, we propose a task-independent graph representation model, called as graph deconvolutional network (GDN), and corresponding unsupervised learning algorithm in this paper. Different with graph convolution network (GCN) from the scratch, which produces embeddings by convolving input attribute vectors with learned filters, the embeddings of the proposed GDN model are desired to be convolved with filters so that reconstruct the input node attribute vectors as far as possible. The embeddings and filters are alternatively optimized in the learning procedure. The correctness of the proposed GDN model is verified by multiple tasks over several datasets. The experimental results show that the GDN model outperforms existing alternatives with a big margin.