DeepGRAND: Deep Graph Neural Diffusion

Graph neural networks (GNNs) have achieved remarkable success in numerous domains. Nevertheless, many popular GNNs are known to suffer from over-smoothing. This phenomenon causes node features to become indistinguishable and hurts the classification accuracy as layer depth increases, which limits their effectiveness at capturing complex and long range graph interactions. We propose the Deep Graph Neural Diffusion (DeepGRAND), a continuous-depth graph neural network that is based on the diffusion process on graphs that theoretically alleviates the over-smoothing issue. DeepGRand pertubes the learnable graph diffusivity and re-scales the underlying diffusion equation by a data-dependent term. We empirically show that DeepGRAND mitigates the accuracy drop-off caused by over-smoothing and surpasses the best accuracy achieved by popular GNNs on various graph deep learning benchmarks. We further demonstrate the advantage of DeepGRAND over many existing graph neural networks in the low label rate regimes.