Towards Deeper Graph Neural Networks via Layer-Adaptive

Graph neural networks have achieved state-of-the-art performance on graph-related tasks. Previous methods observed that GNNs' performance degrades as the number of layers increases and attributed this phenomenon to over-smoothing caused by the stacked propagation. However, we proved experimentally and theoretically that it is overfitting rather than propagation that causes performance degradation. We propose a novel framework: layer-adaptive GNN (LAGNN) consisting of two modules: adaptive layer selection and random Droplayer, which can adaptively determine the number of layers and thus alleviate overfitting. We attached this general framework to two representative GNNs and achieved consistency improvements on six representative datasets.