I want to create Graph Neural Networks that can deal with heterogenous graphs (1 node type, multiple edge types) that allow multiple edges between 2 nodes.
For example, I have a set of similar graphs that look like the following figure:
where each node has an attribute vector, each edge has an attribute vector and an edge type. Given a directed graph with black connections only, the learned model has to predict all the remaining links (green, red, blue) as precisely as possible. For each pair of node, we allow at most 1 green, 1 red, 1 blue and 1 black with arbitrary direction (i.e. A->B or B-A, or self-loops).
I'm thinking of the adjacency reconstruction task, but the challenges are:
- Create such dataset/find similar datasets to experiment on
- Handle such complex graphs with GNNs as graph encoder.
My initial idea would be utilizing Graph Autoencoder and Graph Variational Autoencoder for such task, and expand Graph Attention Networks (GAT) as the Graph Encoder since it can already handle edge features by design. Could you give me some pointers how I could model such dataset and GNNs for the task? Thank you.