Towards Semi-Supervised Universal Graph Classification

Abstract

Graph neural networks have pushed state-of-the-arts in graph classifications\nrecently. Typically, these methods are studied within the context of supervised\nend-to-end training, which necessities copious task-specific labels. However,\nin real-world circumstances, labeled data could be limited, and there could be\na massive corpus of unlabeled data, even from unknown classes as a\ncomplementary. Towards this end, we study the problem of semi-supervised\nuniversal graph classification, which not only identifies graph samples which\ndo not belong to known classes, but also classifies the remaining samples into\ntheir respective classes. This problem is challenging due to a severe lack of\nlabels and potential class shifts. In this paper, we propose a novel graph\nneural network framework named UGNN, which makes the best of unlabeled data\nfrom the subgraph perspective. To tackle class shifts, we estimate the\ncertainty of unlabeled graphs using multiple subgraphs, which facilities the\ndiscovery of unlabeled data from unknown categories. Moreover, we construct\nsemantic prototypes in the embedding space for both known and unknown\ncategories and utilize posterior prototype assignments inferred from the\nSinkhorn-Knopp algorithm to learn from abundant unlabeled graphs across\ndifferent subgraph views. Extensive experiments on six datasets verify the\neffectiveness of UGNN in different settings.\n