Semi-supervised learning with graph embeddings
Python
Switch branches/tags
Nothing to show
Clone or download
Latest commit af6abf4 Sep 28, 2016
Permalink
Failed to load latest commit information.
data add ally Jul 10, 2016
LICENSE add license Jun 28, 2016
README.md Update README.md Sep 28, 2016
base_model.py init Jun 26, 2016
ind_model.py init Jun 26, 2016
layers.py init Jun 26, 2016
test_ind.py dir name Jun 26, 2016
test_trans.py dir name Jun 26, 2016
trans_model.py init Jun 26, 2016

README.md

Planetoid

Introduction

This is an implementation of Planetoid, a graph-based semi-supervised learning method proposed in the following paper:

Revisiting Semi-Supervised Learning with Graph Embeddings. Zhilin Yang, William W. Cohen, Ruslan Salakhutdinov. ICML 2016.

Please cite the above paper if you use the datasets or code in this repo.

Run the demo

We include the Citeseer dataset in the directory data, where the data structures needed are pickled.

To run the transductive version,

python test_trans.py

To run the inductive version,

python test_ind.py

You can refer to test_trans.py and test_ind.py for example usages of our model.

Models

The models are implemented mainly in trans_model.py (transductive) and ind_model.py (inductive), with inheritance from base_model.py. You might refer to the source files for detailed API documentation.

Prepare the data

Transductive learning

The input to the transductive model contains:

  • x, the feature vectors of the training instances,
  • y, the one-hot labels of the training instances,
  • graph, a dict in the format {index: [index_of_neighbor_nodes]}, where the neighbor nodes are organized as a list. The current version only supports binary graphs.

Let L be the number of training instances. The indices in graph from 0 to L - 1 must correspond to the training instances, with the same order as in x.

Inductive learning

The input to the inductive model contains:

  • x, the feature vectors of the labeled training instances,
  • y, the one-hot labels of the labeled training instances,
  • allx, the feature vectors of both labeled and unlabeled training instances (a superset of x),
  • graph, a dict in the format {index: [index_of_neighbor_nodes]}.

Let n be the number of both labeled and unlabeled training instances. These n instances should be indexed from 0 to n - 1 in graph with the same order as in allx.

Preprocessed datasets

Datasets for Citeseet, Cora, and Pubmed are available in the directory data, in a preprocessed format stored as numpy/scipy files.

The dataset for DIEL is available at http://www.cs.cmu.edu/~lbing/data/emnlp-15-diel/emnlp-15-diel.tar.gz. We also provide a much more succinct version of the dataset that only contains necessary files and some (not very well-organized) pre-processing code here at http://cs.cmu.edu/~zhiliny/data/diel_data.tar.gz.

The NELL dataset can be found here at http://www.cs.cmu.edu/~zhiliny/data/nell_data.tar.gz.

In addition to x, y, allx, and graph as described above, the preprocessed datasets also include:

  • tx, the feature vectors of the test instances,
  • ty, the one-hot labels of the test instances,
  • test.index, the indices of test instances in graph, for the inductive setting,
  • ally, the labels for instances in allx.

The indices of test instances in graph for the transductive setting are from #x to #x + #tx - 1, with the same order as in tx.

You can use cPickle.load(open(filename)) to load the numpy/scipy objects x, y, tx, ty, allx, ally, and graph. test.index is stored as a text file.

Hyper-parameter tuning

Refer to test_ind.py and test_trans.py for the definition of different hyper-parameters (passed as arguments). Hyper-parameters are tuned by randomly shuffle the training/test split (i.e., randomly shuffling the indices in x, y, tx, ty, and graph). For the DIEL dataset, we tune the hyper-parameters on one of the ten runs, and then keep the same hyper-parameters for all the ten runs.