Addressing the Limitations of Graph Convolutional Networks

Addressing the Limitations of Graph Convolutional Networks

Explore the challenges and potential solutions for Graph Convolutional Networks (GCNs) in this 28-minute DS4DM Coffee Talk presented by Sitao Luan from McGill University. Delve into the world of graph-structured problems and learn how GCN-based approaches have made significant strides in solving complex issues. Examine the limitations of GCNs, including the over-smoothing problem that hinders deep GCNs from fully utilizing multi-scale information, and the heterophily problem that causes graph-aware models to underperform compared to graph-agnostic models. Gain insights into proposed methods for addressing these challenges and discover current research directions in the field of graph convolutional neural networks.