Two-Stage Training of Graph Neural Networks for Graph Classification

Graph neural networks (GNNs) have received massive attention in the field of machine learning on graphs. Inspired by success networks, a line research has been conducted to train GNNs deal with various tasks, such as node classification, graph and link prediction. In this work, our task interest is classification. Several GNN models proposed shown great accuracy task. However, question whether usual training methods fully realize capacity models. we propose two-stage framework based triplet loss. first stage, trained map each Euclidean-space vector so that graphs same class are close while those different classes mapped far apart. Once well-separated labels, classifier distinguish between classes. This method generic sense it compatible any model. By adapting five method, demonstrate consistent improvement utilization GNN’s allocated over original model up $$5.4\%$$ points 12 datasets.