As Graph Neural Networks (GNNs) has become increasingly popular, there is a wide interest of designing deeper GNN architecture. However, deep GNNs suffer from the oversmoothing issue where the learnt node representations quickly become indistinguishable with more layers. This blog features a simple yet effective technique to build a deep GNN without the concern of oversmoothing. The new architecture, Network in Graph Neural Networks (NGNN) inspired by the network-in-network architecture for computer vision, has shown superior performance on multiple Open Graph Benchmark (OGB) leaderboards.
This is a companion discussion topic for the original entry at https://www.dgl.ai/blog/2022/11/28/ngnn.html