This document presents the Geom-GCN model for graph neural networks. Geom-GCN maps graphs to continuous latent spaces to build structural neighborhoods for message passing aggregation. It uses a bi-level aggregator operating on these structural neighborhoods to update node representations while maintaining permutation invariance. Geom-GCN designs geometric relationships in Euclidean and hyperbolic spaces to define neighborhoods. It achieves state-of-the-art performance on benchmarks while addressing limitations of existing message passing approaches in capturing structure and long-range dependencies.