Schziophrenia PPI network
$A$: adjacency matrix, $X$: input node features, $l$: node/edge labels
Graph convolution of a input $x$ and a filter $g \in \mathbb{R}^N$ is defined as
$$ x *_G g = \mathcal{F}^{-1}(\mathcal{F}(x) \odot \mathcal{F}(g)) = U(U^Tx \odot U^Tg) $$For efficiency, define filters as Chebyshev polynomials.
Apply the same graph convolution layer to update hidden representations.
Apply different graph convolution layers to update hidden representations.
$h_v^t$: vertex features, $e_{uv}$: edge features, $N(v)$: neighbors of vertex $v$
Reducing number of vertecies of a graph.
diffusion convolution + sequence to sequence architecture + scheduled sampling technique
diffusion convolution + sequence to sequence architecture + scheduled sampling technique