konstantinklemmer/pe-gnn

PyTorch implementation of the paper "Positional Encoder Graph Neural Networks for Geographic Data"

PyTorch implementation of PE-GNN

(Architecture of a naive GCN versus that of PE-GNN, enhanced with a positional encoder.)

This is the official repository for the AISTATS 2023 paper Positional Encoder Graph Neural Networks for Geographic Data (Konstantin Klemmer, Nathan Safir, Daniel B. Neill).

Structure

The source code for PE-GNN (using PyTorch) can be found in the src folder. Its built on PyTorch Geometric (ICLR-W, 2019) and Space2Vec (ICLR, 2020).

We also provide an interactive example notebook to test PE-GNN via Google Colab

Citation

If you want to cite our work, you can use the following reference:

@InProceedings{pmlr-v206-klemmer23a,
  title = 	 {Positional Encoder Graph Neural Networks for Geographic Data},
  author =       {Klemmer, Konstantin and Safir, Nathan S. and Neill, Daniel B.},
  booktitle = 	 {Proceedings of The 26th International Conference on Artificial Intelligence and Statistics},
  pages = 	 {1379--1389},
  year = 	 {2023},
  editor = 	 {Ruiz, Francisco and Dy, Jennifer and van de Meent, Jan-Willem},
  volume = 	 {206},
  series = 	 {Proceedings of Machine Learning Research},
  month = 	 {25--27 Apr},
  publisher =    {PMLR},
  pdf = 	 {https://proceedings.mlr.press/v206/klemmer23a/klemmer23a.pdf},
  url = 	 {https://proceedings.mlr.press/v206/klemmer23a.html},
}

pe-gnn's Issues

How to do spatial interpolation ?

Hi thanks for sharing great works!

PE-GCN paper shows performance on 1) Spatial interpolation, 2) Spatial regression and spatial interpolation predicts outcomes from spatial coordinate alone (without any node feature)

• I wonder how I can do spatial interpolation on this repo.
• It seems that x (node feature) assumed to be always exist.

  • pe-gnn/src/model.py

    Line 402 in 58f43ff

    def forward(self, x, c, ei, ew):

Thanks!

AttributeError: 'NoneType' object has no attribute 'knn_graph'

Looks like there is an issue with constructing the graph in the code. This error comes regardless of which dataset is being used

AttributeError

Thank you for your interesting paper!

I attempted to run your code, but unfortunately, an error occurred with the message "AttributeError: 'NoneType' object has no attribute 'knn_graph'". I have already checked the loading of the datasets, it completes correctly.

Maybe you have an idea how to fix this error?

Problems with function lw_tensor_local_moran

When trying to train with the auxiliary task, the following error appears for the execution of lw_tensor_local_moran() function:
RuntimeError: Can't call numpy() on Tensor that requires grad. Use tensor.detach().numpy() instead.

Changing the function from:
zl = torch.tensor(w_sparse * z).to(device)

To:
zl = torch.tensor(w_sparse * z.detach()).to(device)

Solved for me.

How to run GraphSAGE?

The Colab code mentions that "layers like GAT or GraphSAGE can be used interchangeably with the GCN layers". However, when I replace GCNConv with the SAGEConv function, it throws an index error—debugging shows that the error is in the LossWrapper class of model.py. Please let me know how one can run models like GraphSAGE and GAT.