Labeled Convolutions

Implementation of Graph Convolution Layer for graphs with nodes of two different types.

Usage

The code provides LabeledConv class, which can be directly used instead of the usual GCNConv layer in your Graph Neural Network. Instead of

y = GCNConv(x, edge_data)

use the line

y = LabeledConv(x, node_type0, node_type1, edge_data00, edge_data01, edge_data10, edge_data11)

To make LabeledConv flexible, the implementation asks to provide vectors of 1's identifying nodes of two types, as well as edge data.

Experiments

Subgraph properties

In some data we have to determine properties of subgraph $S$ of a given graph $G$. It follows that there are nodes of two types: nodes from subgraph $S$ and nodes not from $S$.

Minimal Vertex Cover

During a decision process on a graph, some nodes are selected. This turns their 1-hop neighbours into covered nodes. In each step of the decision process there are nodes of two types: covered / not covered.