One GCN layer, step by step

This lab applies the rule H' = ReLU( D⁻½ Ã D⁻½ H W ) to the four-vertex cycle graph used in the article. Edit the input features H and weights W, then watch every intermediate matrix update. Begin by reproducing the example; next, set one feature of a vertex to zero and observe how it reappears, diluted, among its neighbors after one layer.

Input H₀ (4×2)
Weights W (2×2)
Adjacency A (fixed)
1· Ã = A + I
2· degrees (with self-loops)
3· Â = D⁻½ Ã D⁻½
4· propagation  H₀
5· transformation (Â H₀) W
6· H₁ = ReLU(·)