zjzijielu / gnn-positional-structural-node-features

Official repository of "On Positional and Structural Node Features for Graph Neural Networks on Non-attributed Graphs", CIKM 2022

On Positional and Structural Node Features for Graph Neural Networks on Non-attributed Graphs (CIKM 2022)

Summary

This repository is the official implementation of the paper. If you would like to use/modify the code or reproduce the experiment results, please remember to cite this paper:

Hejie Cui, Zijie Lu, Pan Li, Carl Yang: On Positional and Structural Node Features for Graph Neural Networks on Non-attributed Graphs. Proceedings of the International Conference on Information and Knowledge Management (CIKM) 2022

System and software config

The experiments are run locally on Ubuntu 20.04, with NVCC 10.2 and Python 3.8.5. The Python module versions can be found in requirements.txt.

Requirements

To run experiments, first set up virtualenv. After activating the virtul environment, run

pip3 install -r requirements.txt

Positional and structural node classification

RUN

cd graphsage-simple
bash train_nodes.sh

You can specify the initialization method, dataset, epoch, feature dimension and learning rate in train_nodes.sh. The results can be found in results folder.

Graph classification

Refer to gnn_comparison/experiment.sh for example runbook.

Deepwalk

Deepwalk is one of the initialization methods that we explore in this project. To generate deepwalk features, refer to https://github.com/phanein/deepwalk.

Make sure to include dataset directory in the --input and --ouput flag value, and set the --format.

gnn-positional-structural-node-features's Issues

Missing "--use-eigen" feature initialization code

got it, closed.

about random and one-hot initialization

I saw that in your model.py code, for the one-hot initializer you set the input feature embedding to be learnable, but for random normal initializer you set require_grad to false. It seems for the one-hot setting the feature should be not learnable but for the random one it should be trainable. Do I understand it correctly?