SEANO: Semi-supervised Embedding in Attributed Networks with Outliers

SEANO provides an effective method for graph embedding on attributed networks that contain outliers. For more technical details, refer to our paper. Here we publish our code and dataset along with a brief instruction on how to run the code.

Code and data

They can be downloaded from here.

Required packages

  • tensorflow
  • numpy
  • scipy

How to run

Use python main.py to run the code with the default parameters. For more details of the arguments, see main.py. You might need to use the validation dataset to tune some hyper-parameters (e.g., graph_context_batch_size, label_context_batch_size, etc.) for your own dataset.

Inputs

The list of required input files is shown below. The node index should be continuous starting from 0 and all the input files are stored in cPickle format. xxx means the name of the dataset, e.g., cora.

  • xxx.X: attributes of all nodes (sorted by id), e.g., cora.X.
  • xxx.Y: label vectors of all nodes, e.g., cora.Y.
  • xxx.graph: graph stored in dictionary format. node id –> list of neighbors.
  • xxx.labeled.index: labeled nodes for semi-supervised training.
  • xxx.val.index: index of validation nodes.
  • xxx.test.index: index of testing set. In the inductive case, this part will be excluded from training.
  • xxx.rnd_walks.npy (optional): random walk files. If missing, new random walks will be generated and written to this file.

Outputs

All the output files are located in ./output/, all of which are stored in cPickle format. There are three outputs:

  • xxx.embed: Embedding of each node in the graph.
  • xxx.pred_label: Predicted label of each node.
  • xxx.outlier_score: Outlier score of each node.

Note: if inductive learning is enabled through --inductive, the prefix of the outputs would be xxx_ind..

Reference

J. Liang, P. Jacobs, J. Sun, S. Parthasarathy. “Semi-supervised Embedding in Attributed Networks with Outliers”. To appear in Proceedings of SIAM International Conference on Data Mining (SDM’18), 2018. [PDF][bib]