- Tuyet Anh Dao
- 0 Comments
- 480 Views
AISTATS 2024
Abstract:
We propose the Kuramoto Graph Neural Network (KuramotoGNN), a novel class of continuous-depth graph neural networks that employ the Kuramoto model, renowned for analyzing synchronization in coupled oscillator systems. KuramotoGNN mitigate the over-smoothing phenomenon, in which node features in GNNs become indistinguishable as the number of layers increases. Additionally, our work is the pioneer in highlighting the theoretical connection between over-smoothing and synchronization. This connection offers valuable insights into the stability of the model and sheds light on the behavior of GNNs.
Paper: https://arxiv.org/abs/2311.03260
Code: TBU