Learning graph-level representation from local-structural distribution with Graph Neural Networks

Abstract

Graph Neural Networks has been proved to be successful in learning the latent vector representation for local structures. While applying this technique to some real-world tasks, such as property prediction for molecules, a graph-level representation generator should be required to integrate all these local-structural embeddings. Limited by varying graph scale and the absence of node order, existing graph-level representation generation approaches usually adopt rough ways, e.g., compressing all local-structural embeddings into a single vector, which cannot simultaneously retain both the local- and global-structural information in graph-level representations. To address this problem, this paper introduces a novel and more powerful approach to learn graph-level representation from local-structural distribution. Firstly, our approach employs a batch strategy to discretize the embeddings of local structures by their structural semantics, which can provide interactive information to approximately align the local-structural embeddings for varying graphs. According to Wasserstein metric, the aligned structural information is beneficial to capture the similarity among graphs. Then, our approach further integrates these aligned local-structural embeddings to construct a resolution-controllable structural histogram as the graph-level representation. Without over-compression, more local-structural information can be preserved. The experimental results show that our approach substantially outperformed the baselines on a range of real-world datasets in both graph classification and regression tasks.