This is a TensorFlow implementation of Improved Graph Convolutional Networks (IGCN) and Generalized Label Propagation (GLP) for the task of (semi-supervised) node classification, as described in our paper:

Li, Qimai, Xiao-Ming Wu, Han Liu, Xiaotong Zhang, and Zhichao Guan. "Label Efficient Semi-Supervised Learning via Graph Filtering." In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2019.

• python3
• tensorflow (1.5.0 and 1.9.0 are tested)
• networkx
• numpy
• scipy

Anaconda environment is recommended.

python train.py -v

Reproduction may require average accuracy over multiple runs by adding option --repeat 50.

# citation network, 20 labels per class
python train.py --pset config_citation.large_label_set --dataset cora
python train.py --pset config_citation.large_label_set --dataset citeseer 
python train.py --pset config_citation.large_label_set --dataset pubmed
python train.py --pset config_citation.large_label_set --dataset large_cora --layer-size '[64]'

# citation network, 4 labels per class
python train.py --pset config_citation.small_label_set --dataset cora
python train.py --pset config_citation.small_label_set --dataset citeseer
python train.py --pset config_citation.small_label_set --dataset pubmed
python train.py --pset config_citation.small_label_set --dataset large_cora --layer-size '[64]'

# NELL
python train.py --pset config_nell.large_label_set --dataset nell.0.1
python train.py --pset config_nell.small_label_set --dataset nell.0.01
python train.py --pset config_nell.small_label_set --dataset nell.0.001

In order to use your own data, you have to provide

• an N by N adjacency matrix (N is the number of nodes),
• an N by D feature matrix (D is the number of features per node), and
• an N by E binary label matrix (E is the number of classes).

Have a look at the load_data() function in gcn/utils.py for an example.

In this example, we load citation network data (Cora, CiteSeer, PubMed, Large Cora and NELL). The original datasets of Cora, CiteSeer, PubMed can be found here: https://linqs.soe.ucsc.edu/data. NELL's official sit is here: http://rtw.ml.cmu.edu/.

In our version (see data folder) Cora, CiteSeer, PubMed and NELL are split by [3].

Large Cora is constructed from Andrew McCallum's Cora Project, and processed by our own. If you use Large Cora dataset, please cite our paper "Label Efficient Semi-Supervised Learning via Graph Filtering" and McCallum's paper "Automating the Construction of Internet Portals with Machine Learning"

You can choose between the following models:

• IGCN: Improved Graph Convolutional Networks [1]
• GLP: Generalized Label Propagation [1]
• GCN: Graph convolutional network [2]
• LP: Label Propagation [4]
• MLP: Basic Multi-Layer Perceptron that supports sparse inputs.

Please cite our paper if you use this code or Large Cora dataset in your own work:

@inproceedings{li2019label,
  title={Label Efficient Semi-Supervised Learning via Graph Filtering},
  author={Qimai Li and Xiao-Ming Wu and Han Liu and Xiaotong Zhang and Zhichao Guan},
  booktitle={Conference on Computer Vision and Pattern Recognition},
  year={2019},
  url={https://arxiv.org/abs/1901.09993},
}

Thanks for Kipf's implementation of GCN, on which this repository is initially based.

[1.] Q. Li, X.-M. Wu, H. Liu, X. Zhang, and Z. Guan. Label efficient semi-supervised learning via graph filtering. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019.

[2.] T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. In International Confer- ence on Learning Representations, 2017.

[3.] Z. Yang, W. W. Cohen, and R. Salakhutdinov. Revisiting semi-supervised learning with graph embeddings. In International Conference on Machine Learning, pages 40–48, 2016.

[4.] X. Wu, Z. Li, A. M. So, J. Wright, and S.-f. Chang. Learning with Partially Absorbing Random Walks. In Conference on Neural Information Processing Systems, pages 3077–3085, 2012.