Sample Code for Graph Partition Neural Networks
Switch branches/tags
Nothing to show
Clone or download
Latest commit 7e4f929 Mar 26, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
config Initial commit Nov 15, 2017
data Initial commit Nov 15, 2017
gpnn Initial commit Nov 15, 2017
.gitignore Initial commit Nov 15, 2017
LICENSE Initial commit Nov 15, 2017 Add link of paper Mar 26, 2018 Initial commit Nov 15, 2017 Initial commit Nov 15, 2017 Initial commit Nov 15, 2017 Initial commit Nov 15, 2017


This is the Tensorflow implementation of Graph Partition Neural Networks as described in the following paper:

  title={Graph Partition Neural Networks for Semi-Supervised Classification},
  author={Renjie Liao and Marc Brockschmidt and Daniel Tarlow and Alexander~L. Gaunt and Raquel Urtasun and Richard Zemel},
  journal={arXiv preprint arXiv:1803.06272},


To set up experiments of NELL/DIEL, please run the following scripts:


Note that since these two datasets are large, it may take a while to finish. You may also need to switch to CPU mode before running demos by:



tensorflow(>= 1.0), numpy, scipy, sklearn

Run Demos

  • To run experiments X where X is one of {citeseer, cora, pubmed, nell, diel}:

    python -c config/gpnn_X.json

  • For experiments on NELL, you can specify the label rate by changing label_rate of config/gpnn_nell.json

  • For experiments on DIEL, you can specify the split id by changing split_id of config/gpnn_diel.json


We list some notable hyper-parameters here and you can refer to configuration files under the config folder for more details.

  • num_pass: # Propagation pass
  • num_cluster: # Clusters
  • prop_step_intra: # Propagation step within cluster
  • prop_step_inter: # Propagation step between clusters
  • decomp_method: one of {"spectral_cluster", "multi_seed_flood_fill"}
  • hidden_dim: dimension of state vector
  • aggregate: aggregation method, one of {"avg", "sum", "min", "max"}
  • msg_type: message function type, one of {"msg_embedding", "msg_mlp"}
  • update_type: update function type, one of {"GRU", "MLP"}
  • update_MLP_dim: hidden dimension of update MLP
  • update_MLP_act: activation function of update MLP, one of {"relu", "tanh", "sigmoid", null}
  • output_MLP_dim: similar to update_MLP_dim
  • output_MLP_act: similar to update_MLP_act


  • An example of specifying a 2 hidden layer output MLP with "tanh" as activation function:
    output_MLP_dim = [128, 128], output_MLP_act = ["tanh", "tanh"].
  • We also provide an easy-to-use implementation of LSTM in which could potentially be used as update function. We will support this feature soon.
  • By setting num_cluster = 1 and update_type = "GRU", the resultant model is roughly the same as Gated Graph Neural Network except the input and output models are slightly different from the ones described in the original paper.

Customized Usage

To use our code for your customized problem, you need to prepare the following (pickle) files:

  • your_dataset.graph: graph, python dictionary, key = node id, value = list of neighbor ids, id ranges from 0 to N-1.
  • your_dataset.feature: feature, N by D numpy array (D is the number of features per node).
  • your_dataset.label: label, N numpy array (integer label within range(C), C is the number of classes).
  • your_dataset.split: mask of split, N numpy array (0, 1, 2 stand for train/validation/test respectively).
  • Append the customized dataset information in config/dataset_info.json by setting label_size, feat_dim, num_nodes, num_valid, data_folder.

Please refer to gpnn/reader/ for more information.


Please cite our paper if you use this code in your research work.


Please submit a Github issue or contact if you have any questions or find any bugs.


This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit

When you submit a pull request, a CLA-bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact with any additional questions or comments.