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

Abstract

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