This is the accompany python package for the ICML 2019 paper Learning Discrete Structures for Graph Neural Networks
It implements the method LDS and its variant KNN-LDS and reproduces experiments reported in the paper.
The code is written Python 3.6 and TensorFlow version 1 (tested on versions 1.12 and 1.16). It requires scikit-learn >= 0.21.2 and the python packages
UCI datasets should be loaded automatically, while graph-based datasets (Cora and Citeseer) are included in the GCN package, available here. Place the relevant files in the folder lds/data.
FMA dataset (we used the small version) should also be downloaded, please email the authors if interested.
python setup.py install
The scripts contained in lds.py should work also without installing the package.
Navigate to lds_gnn folder.
The main script is in the file lds.py. The options are
-d: the evaluation dataset. Available datasets are iris, wine, breast_cancer, digits, 20newstrain,
20news10, cora, citeseer, fma. Default breast_cancer
-m: the method: lds or knnlds. Default knnlds
-s: the random seed. Default 1
-e: the percentage of missing edges (valid only for cora and citeseer dataset). Default 50
For experiments with incomplete graphs on Cora and Citeseer, run
python lds.py -m lds -e {an integer between 0 and 100} -d {cora or citeseer} -s {if you want to specify random seed}
For experiments in semi-supervised learning (with no input graph), run
python lds.py -m knnlds -d {any available dataset} -s {if you want to specify random seed}
The code will run a small grid search to select some method's parameters such as the
(outer) optimization learning rate and the number of truncation steps to compute the hypergradeient. It
will output the test accuracy of the best found model, according to the ''early stopping accuracy''.
It will also create one log file per single experiment in the folder lds/results,
which can be successively loaded (e.g. in a notebook) with the function lds.load_results()
for inspection and visualization .
Please note that the package does not include implementations of baseline methods.
Please take a look at LICENCE.txt
If you use this package, please cite
@InProceedings{franceschi2019learning,
title = {Learning Discrete Structures for Graph Neural Networks},
author = {Luca Franceschi and Mathias Niepert and Massimiliano Pontil and Xiao He},
booktitle = {Proceedings of the 36th International Conference on Machine Learning},
year = {2019}
}