Exploring Structural Feature Correlations via Graph Neural Networks (version 1.0.0)

function: give correlations between each two structural features, or between one and leftover features.

Brief: torch based program for graph feature to feature predictions.

This project is supervised by Zhitao (Rex) Ying from Stanford. Paper has been admitted by ECML PKDD Graph Embedding and Mining(GEM) workshop. Paper will be published in the proceedings of ECML PKDD workshop sooner or later.

Installation

please ensure that torch has been installed successfully in your computer, check with:

$ python -c "import torch; print(torch.__version__)"
>>> 1.8.0

please install torch_geometric with correct torch version and cuda version. In our experiment environment, torch version is 1.8.0 and cuda version is 11.1 for both GPU(GTX 2060Super) and torch. (It's important since different versions will cause imcompatible situations)

pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-sparse -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-cluster -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html
pip install torch-geometric

please install networkx. this package is used for calculating graph features and generating synthetic graphs.

pip install networkx

pipeline model

Dataset(Initial)

1. node classification datasets:

Cora CiteSeer PubMed

2. graph classifcation datasets:

ENZYMES PROTEINS NCI1

3. Synthetic datasets

Geometric synthetic datasets are generated by networkx.

Procedures

1. Generating graph structural features

Graph features have been calculated already, where you can find it in \Result folder. {k}_property.txt, where k is the name of each dataset. Please make sure that networkx is installed. If you want to generate them by youself, run:

python generate_property_planetoid.py ## for planetoid dataset
python generate_proprty_tuda.py ## for tudataset

under property_process folder. Functional file is graph_property.py, where you can add more structural features to the given list.

2. Single feature to feature prediction (Fea2Fea-single)

For single feature to single feature prediction, go to the src folder and run f_f_Citation.py if you want to run experiments on Planetoid datasets. If you want to perform experiments on TUDatasets, then please run f_f_TU.py

   python f_f_Citation.py 
   python f_f_TU.py

You need to correct the path at the moment before it's been corrected to the right relative path. Results are saved in Results/{dataset}.

3. Multiple features to single feature prediction (Fea2Fea-multiple)

What's different from Fea2Fea-single is that predicted objective should be set in advance in the parameter list which should be passed to the command line. For example: run

   python aug_Citation.py --dataset=Cora --aim_feature = 2 ## prediction on Cora dataset and predict clustering coefficient
   python aug_TU.py --dataset=ENZYMES --aim_feature = 4 ## prediction on ENZYMES dataset and predict average path length (shortest)

The predicted objective will be filtered by function all_possible_concatenation firstly, generating the all possible concatenation group in the array ans. After that, it will go through graph neural network models to reach prediction accuracy for each element in the array and record average acc. The error bar-plot will be shown on the screen (for each dimension).

4. Real-world applications (important)

After analyzing the potential irredundant feature groups, we are going to concatenate them with initial graph features to make node or graph classifications. run:

   python citation_realworld.py --dataset=Cora --graphconv=GIN --o.concat_method=SimpleConcat
   ## prediction on Cora dataset, using GIN model in embedding layer and just simple concatenate each augmented structural feature.

Finally, you will reach acc for simple/bilinear/NTN + number of input features. This is a small trick where we only randomly choose from the array. You can choose the best one among all results. You can add your choice in the file, set the ans[0] to what you'd want the input feature idxes be.

5. Hyper-parameter tests

In this part, we introduce the hyper-parameter tests.

    python hyper_param.py [option list]

Parameter list:

Parameters	Type	Default Value	Description
input_feature	int	0	input feature index
aim_feature	int	1	output feature index
task	str	`node`	node or graph dataset
dataset	str	`Cora`	dataset name
hyperparameter	str	`binning`	binning or depth or threshold tunning
min_bins	int	2	minimum number of bins
max_bins	int	6	maximum number of bins
min_depth	int	2	minimum depth of GNN architecture
max_depth	int	6	maximum depth of GNN architecture
hidden_dim	int	2	hidden dimension
batchnorm	bool	0	if BatchNorm
embedding	str	`GIN`	graph embedding method
threshold	float	0.8	threshold for filtering irredundant features

Results will be shown on the screen. By the way, you can only perform hyperparameter once, choosing from binning, depth or threshold.

visualization

input feature distribution

If you want to draw distribution to check whether the number of bins is reasonable, run

   python plot_dist.py # under src folder

jupyter notebook playground (for reference)

to compare concatenation methods: different_concat_method_node.ipynb

to explore different embeddings: explore_embedding.ipynb

Most importantly, in the future, we are going to add more features and more graph neural network models to ensure the model's robustness.

Citation

If you use to cite Fea2Fea in your research paper, please consider citing:

@misc{xie2021fea2fea,
      title={Fea2Fea: Exploring Structural Feature Correlations via Graph Neural Networks}, 
      author={Jiaqing Xie and Rex Ying},
      year={2021},
      eprint={2106.13061},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Currently its not been published by Springer, but will be in a few months.