Layer formula

H^(l+1) = σ(D̂^(-1/2)  D̂^(-1/2) H^l W^l). Where  = A + I (self-loops), D̂ = degree matrix of Â.

Advertisement

Interpretation

Each layer: aggregate neighbor features (mean-like), transform via W, apply nonlinearity. Two layers ≈ 2-hop neighborhood info.

Advertisement

Semi-supervised classification

Some labeled, rest unlabeled. Cross-entropy on labeled. Softmax over classes.

Limitations

Doesn't distinguish importance of neighbors. Over-smoothing at depth. Fixed graph structure needed.