The Third Common Interface for Graph Neural Networks

Although graph neural networks (GNNs) are powerful tools for analyzing data that can be represented as a graph, they generally suffer from the problem of over-smoothing as the number of layers in the networks increases. This paper focuses on two recent methods for overcoming this problem, namely JKNet-lstm and TwinGNN, and presents a simpler GNN interface that achieves a similar reduction in over-smoothing and its implementations that require much lower computational cost. The proposed approach utilizes the relationship between the convolution operation in GNNs and the reproductive property of the normal distribution. In particular, our method does not require any computationally intensive long short-term memory calculations, and achieves complexity that is independent of the number of layers in the GNN. We present mathematical arguments to prove that our proposed interface achieves the same functionality as JKNet-lstm and TwinGNN, and perform experiments on six benchmark datasets to demonstrate the remarkable performance of our method in terms of the classification accuracy, computational cost, and degree of over-smoothing. Overall, the proposed interface enables complex graph network structures to be analyzed at reasonable computational cost with no loss of accuracy.