Graph Neural Network: A First Glance

@[TOC]GNN

Resources

從圖(Graph)到圖卷積(Graph Convolution)：漫談圖神經網絡模型 (一)

Vocabulary

• Fixed Point Theorem : a convergency guarantee
• Contraction Map
• BP: Almeida-Pineda vs BPTT

Short Notes

• To make $f$ a Contraction Map: Penalize Jacobian Matrix of $f$ over $H$. I.e. Bound its derivative.
• GNN: stop when converged.
• GNN drawbacks
  • Edges serve only as connections not learned
  • Not suitable for learning Graph Representation: all nodes share info with each other.
• GGNN: replace convergent $f$ with a Gated Unit like in RNN. Use BPTT instead of AP and can output before convergence. Edges now have weights that can be updated.

Q&A

• From Tree to Graph: this is all?
• Spectual Domain vs Spatial Domain
• How to update weights
• Attention