InfoGraph: Unsupervised and Semi-supervised Graph-Level Representation Learning via Mutual Information Maximization

This paper studies learning the representations of whole graphs in both and semi-supervised scenarios. Graph-level representations are in a variety of real-world applications such as predicting the of molecules and community analysis in social networks. Traditional kernel based methods are simple, yet effective for obtaining fixed-length for graphs but they suffer from poor generalization due to-crafted designs. There are also some recent methods based on language (e. g. graph2vec) but they tend to only consider certain substructures (e. g. subtrees) as graph representatives. Inspired by recent progress of representation learning, in this paper we proposed a novel method InfoGraph for learning graph-level representations. We maximize the information between the graph-level representation and the of substructures of different scales (e. g. , nodes, edges, ). By doing so, the graph-level representations encode aspects of the that are shared across different scales of substructures. Furthermore, we propose InfoGraph*, an extension of InfoGraph for semi-supervised. InfoGraph* maximizes the mutual information between unsupervised representations learned by InfoGraph and the representations learned by supervised methods. As a result, the supervised encoder learns from data while preserving the latent semantic space favored by the supervised task. Experimental results on the tasks of graph and molecular property prediction show that InfoGraph is to state-of-the-art baselines and InfoGraph* can achieve performance with state-of-the-art semi-supervised models.