RIGNN: A Rationale Perspective for Semi-Supervised Open-World Graph Classification

Abstract

Graph classification has gained growing attention in the graph machine learning community and a variety of semi-supervised methods have been developed to reduce the high cost of annotation. They usually combine graph neural networks (GNNs) and extensive semi-supervised techniques such as knowledge distillation. However, they adhere to the close-set assumption that unlabeled graphs all belong to known classes, limiting their applications in the real world. This paper goes further, investigating a practical problem of semi-supervised open-world graph classification where these unlabeled graph data could come from unseen classes. A novel approach named Rationale-Informed GNN (RIGNN) is proposed, which takes a rationale view to detect components containing the most information related to the label space and classify unlabeled graphs into a known class or an unseen class. In particular, RIGNN contains a relational detector and a feature extractor to produce effective rationale features, which maximize the mutual information with label information and exhibit sufficient disentanglement with non-rationale elements. Furthermore, we construct a graph-of-graph based on geometrical relationships, which gives instructions on enhancing rationale representations. In virtue of effective rationale representations, we can provide accurate and balanced predictions for unlabeled graphs. An extension is also made to accomplish effective open-set graph classification. We verify our proposed methods on four benchmark datasets in various settings and experimental results reveal the effectiveness of our proposed RIGNN compared with state-of-the-art methods.