Abstract

Designing neural network architectures that can handle data symmetry iscrucial. This is especially important for geometric graphs whose properties areequivariance under Euclidean transformations. Current equivariant graph neuralnetworks (EGNNs), particularly those using message passing, have a limitationin expressive power. Recent high-order graph neural networks can overcome thislimitation, yet they lack equivariance properties, representing a notabledrawback in certain applications in chemistry and physical sciences. In thispaper, we introduce the Clifford Group Equivariant Graph Neural Networks(CG-EGNNs), a novel EGNN that enhances high-order message passing byintegrating high-order local structures in the context of Clifford algebras. Asa key benefit of using Clifford algebras, CG-EGNN can learn functions thatcapture equivariance from positional features. By adopting the high-ordermessage passing mechanism, CG-EGNN gains richer information from neighbors,thus improving model performance. Furthermore, we establish the universalityproperty of the $k$-hop message passing framework, showcasing greaterexpressive power of CG-EGNNs with additional $k$-hop message passing mechanism.We empirically validate that CG-EGNNs outperform previous methods on variousbenchmarks including n-body, CMU motion capture, and MD17, highlighting theireffectiveness in geometric deep learning.