Self-supervised Smoothing Graph Neural Networks

Graph Neural Networks (GNNs) have been leading an effective framework of learning graph representations, and have been demonstrated powerful in numerous tasks [25]. The key of GNNs roots at the repeated aggregation over local neighbors to obtain smoothing node representations, making close nodes have similar representations by filtering out noise existing in the raw node features. Learning GNN models to maintain local smoothness usually depends on supervision signals such as node labels or self-supervision signals extracted from the input graph. Whereas, labels are not always available in many scenarios. Besides, the trained GNNs in supervised and semi-supervised ways are not universal applicable, and only serve the defined learning tasks, e.g., node classification when the GNNs were trained with node labels, or graph classification when the GNNs are trained with graph labels.