PyTorch Geometric High Order Documentation

PygHO is a library for high-order GNN. Ordinary GNNs, like GCN, GIN, GraphSage, all pass messages between nodes and produce node representations. The node representation forms a dense matrix of shape $(n, d)$, where $n$ is the number of nodes and $d$ is the hidden dimension. Existing libraries like PyG can easily implement them.

In constrast, higher-order GNNs (HOGNNs) use node tuples as the message passing unit and produce representations for the tuples. The tuple representation can be of shape $(n, n, d)$, $(n, n, n, d)$, and even more dimensions. Furthermore, to reduce complexity, the representation can be sparse. PyGHO is the first unified library for HOGNNs.

>@inproceedings{PyGHO,
                author = {Xiyuan Wang and Muhan Zhang},
                title = {{PyGHO, a Library for High Order Graph Neural Networks}},
                year = {2023},
}

Notes

Advanced Tutorial

Multiple Tensor