Using nerual network at graph data. The basic formula of this is thta: $H(x)^{(l+1)} = f(H^{(l)},A)$ $f(H(x)^{(l)},A) = \delta(AH^{(l)}W^{(l)})$ $ l $ is the layer of nerual network.

In Thomas Kipf post, we know that GCN works better than ordianry methods, even in three layers and simple active function like ReLU function.

In order to get familiar with GCN, these paper should be finished before next week, and then write short paper review:

• Convolutional Networks on Graphs for Learning Molecular Fingerprints
• Semi-Supervised Classification with Graph Convolutional Networks

And there other related paper focus on other areas:

• Deep Walk
• Trainable Adjancy Matrix

And currently, I still cannot understand why should the biggest eignvalues of adjancy matrix should less than 1 to prove Asymptotic stability. And this stablity often prove the system will keep in low activity.