Learning graph normalization for graph neural networks

Graph Neural Networks (GNNs) have emerged as a useful paradigm to process graph-structured data. Usually, GNNs are stacked multiple layers and node representations in each layer computed through propagating aggregating the neighboring features. To effectively train GNN with layers, normalization techniques necessary. Though existing achieved good results helping training, but they seldom consider structure information of graph. In this paper, we propose two graph-aware techniques, namely adjacency-wise graph-wise normalization, which fully take into account Furthermore, novel approach, termed Attentive Normalization (AGN), learns weighted combination methods, aiming automatically select optimal methods for specific task. We conduct extensive experiments on eleven benchmark datasets, including three single-graph eight multiple-graph experimental provide comprehensive evaluation effectiveness our proposals.