8th International Conference on
Computational Harmonic Analysis

September 12-16, 2022

Ingolstadt, Germany

"Framelet Message Passing: GNNs Propagation with Multiscale Multi-hop Representation and No Oversmoothing"

Wang, Yuguang

Graph neural networks (GNNs) encode adequate feature representations of graphs. The dominant graph feature distillation leverages neural message passing which is a message update method of a specific node feature from one layer to the next by incorporating 1-hop neighbors. It nevertheless usually suffers from oversmoothing. This work proposes Framelet Message Passing that integrates multiscale framelet representation of neighbor nodes in multiple hops in node message update. Our method circumvents oversmoothing with a non-decay Dirichlet energy in propagation. It achieves state-of-the-art performance on real node classification tasks with low computational cost.

« back