Gentle Normalization and Translation in Graph Neural Network for Few-shot Learning

Few-shot learning methods based on graph neural networks (GNNs) have shown powerful capabilities. However, GNNs have a very intractable problem called oversmoothing. The oversmoothing problem refers to that as the number of GNN layers increases, the node information will converge to a similar value, which is difficult to be distinguished, thus reducing the classification performance. In this paper, a Gentle Normalization and Translation (GNT) model is proposed to solve the above problem. On the basis of the original Normalization method, a Gentle Normalization is presented to solve the oversmoothing problem and reduce the model variance by reducing the scaling range. Further, a Translation operation is developed to deal with the oversmoothing problem caused by the ReLU layer. In addition, the Initial Residual is added which can also solve the oversmoothing problem to a certain extent. The experiments on public datasets show that the classification performance has been improved considerably.