Pyramidal Reservoir Graph Neural Network

We propose a deep Graph Neural Network (GNN) model that alternates two types of layers. The first type is inspired by Reservoir Computing (RC) and generates new vertex features iterating non-linear map until it converges to fixed point. second layer implements graph pooling operations, gradually reduce the support features, further improve computational efficiency RC-based GNN. architecture is, therefore, pyramidal. In last layer, remaining vertices are combined into single vector, which represents embedding. Through mathematical derivation introduced in this paper, we show formally how can complexity speed-up convergence dynamical updates features. Our proposed approach design GNNs offers an advantageous principled trade-off between accuracy complexity, extensively demonstrate experiments on large set datasets.