Source code for paper "TriATNE: Tripartite Adversarial Training for Network Embeddings"
The code has been tested under Python 3.7.5, with the following packages installed (along with their dependencies):
- tensorflow == 1.14.0
- gensim == 3.8.0
- networkx == 2.4
- numpy = 1.17.4
- scikit-learn == 0.21.3
- scipy == 1.3.1
- tqdm = 4.39.0 (for displaying the progress bar)
- \data: Store the example inputs.
- \emb: Store the learned embeddings.
- \eval: Store the evaluation scripts for different metric.
- \model: Store the trained models.
- model.py: Define each modules of the model.
- PRE.py: Train-test split for link prediction.
- settings.py: Defines different hyperparameters here.
- TriATNE.py: The main entrance of running.
- utils.py: Codes to load data, split data, prepare batches for training and write embeddings to files.
(1) TriATNE expect an edgelist for the input network, i.e.,
node1 node2 1
node1 node3 1
...
Each row contains two IDs, which are separated by '\t'
(2) Two formats are supported for the input labels. By setting format = 'cora' , we expect a node with multi-labels:
/label1/label2/label3/
/label1/label5/
...
The node ID defaults to line number and starts from 0 to N−1 (N is the number of nodes in the graph)
By setting format = 'blogcat', we expect a node with single label:
node1 label1
node2 label1
...
nodeN label5
The first column is the node’s ID. The second column is the label of corresponding node. They are separated by'\t'.
(3) The output embeddings are as follows:
N d
node1 f1 f2 ... fd
node2 f1 f2 ... fd
...
nodeN f1 f2 ... fd
where N denotes the number of nodes and d denotes the embeddings' dimension. Each subsequent row represents a node.
We take the Citeseer dataset as an exampleand show how to perform classification tasks for TriATNE:
(1) Modify the following parameters in settings.py:
TRAIN_INPUT_FILENAME = '/data/Citeseer_input.txt'
TRAIN_LABEL_FILENAME ='/data/Citeseer_label.txt'
format ='blogcat'APP = 0
(2) Run the code by:
python TriATNE.py
(3) The learned embeddings will be stored in /emb/, and the evaluation results (micro F1 and macro F1) for embeddings will be printed on the screen for each epoch.
We take the wiki-vote dataset as an exampleand show how to perform link prediction tasks for TriATNE:
(1) If you have split the dataset, please move directly to step 2. Otherwise, modify the following parameters in settings.py:
APP = 1
split = True
test_frac = 0.2
FULL_FILENAME = '/data/wiki-vote.txt'
TRAIN_POS_FILENAME = '/data/wiki_train_pos.txt'
TRAIN_NEG_FILENAME = '/data/wiki_train_neg.txt'
TEST_POS_FILENAME = '/data/wiki_test_pos.txt'
TEST_NEG_FILENAME = '/data/wiki_test_neg.txt'
Note that splitting data can take a long time. Therefore, we recommend storing the divided data for later use.
(2) If you have finished parameter setting in step 1, skip to step 3. Otherwise, modify the following parameters in setting.py:
APP = 1
split = False
TRAIN_POS_FILENAME = '/data/wiki_train_pos.txt'
TEST_POS_FILENAME = '/data/wiki_test_pos.txt'
TEST_NEG_FILENAME = '/data/wiki_test_neg.txt'
(3) Run the code by:
python TriATNE.py
(4) The learned embeddings will be stored in /emb/, and the evaluation results (AUC and NDCG@k) for embeddings will be printed on the screen for each epoch.