Code for the paper Asymptotics of L2 Regularized Network Embeddings.
Requires Stellargraph 1.2.1, Tensorflow 2.6.0, scikit-learm 0.24.1, tqdm, along with any other packages required for the above three packages.
To run node classification or link prediction experiments, run
python -m code.train_embed [[args]]
or
python -m code.train_embed_link [[args]]
from the command line respectively, where [[args]] correspond to the
command line arguments
for each function. Note that the scripts expect to run from the parent
directory of the code folder; you will need to change the import
statements in the associated python files if you move them around. The
-h command line argument will display the arguments (with descriptions)
of each of the two files.
| short | long | default | help |
|---|---|---|---|
-h |
--help |
show this help message and exit | |
--dataset |
Cora |
Dataset to perform training on. Available options: Cora,CiteSeer,PubMedDiabetes | |
--emb-size |
128 |
Embedding dimension. Defaults to 128. | |
--reg-weight |
0.0 |
Weight to use for L2 regularization. If norm_reg is True, then reg_weight/num_of_nodes is used instead. | |
--norm-reg |
Boolean for whether to normalize the L2 regularization weight by the number of nodes in the graph. Defaults to false. | ||
--method |
node2vec |
Algorithm to perform training on. Available options: node2vec,GraphSAGE,GCN,DGI | |
--verbose |
1 |
Level of verbosity. Defaults to 1. | |
--epochs |
5 |
Number of epochs through the dataset to be used for training. | |
--optimizer |
Adam |
Optimization algorithm to use for training. | |
--learning-rate |
0.001 |
Learning rate to use for optimization. | |
--batch-size |
64 |
Batch size used for training. | |
--train-split |
[0.01, 0.025, 0.05] |
Percentage(s) to use for the training split when using the learned embeddings for downstream classification tasks. | |
--train-split-num |
25 |
Decides the number of random training/test splits to use for evaluating performance. Defaults to 50. | |
--output-fname |
None |
If not None, saves the hyperparameters and testing results to a .json file with filename given by the argument. | |
--node2vec-p |
1.0 |
Hyperparameter governing probability of returning to source node. | |
--node2vec-q |
1.0 |
Hyperparameter governing probability of moving to a node away from the source node. | |
--node2vec-walk-number |
50 |
Number of walks used to generate a sample for node2vec. | |
--node2vec-walk-length |
5 |
Walk length to use for node2vec. | |
--dgi-sampler |
fullbatch |
Specifies either a fullbatch or a minibatch sampling scheme for DGI. | |
--gcn-activation |
['relu'] |
Determines the activations of each layer within a GCN. Defaults to a single layer with relu activation. | |
--graphSAGE-aggregator |
mean |
Specifies the aggreagtion rule used in GraphSAGE. Defaults to mean pooling. | |
--graphSAGE-nbhd-sizes |
[10, 5] |
Specify multiple neighbourhood sizes for sampling in GraphSAGE. Defaults to [10, 5]. | |
--tensorboard |
If toggles, saves Tensorboard logs for debugging purposes. | ||
--visualize-embeds |
None |
If specified with a directory, saves an image of a TSNE 2D projection of the learned embeddings at the specified directory. | |
--save-spectrum |
None |
If specifies, saves the spectrum of the learned embeddings output by the algorithm. |
| short | long | default | help |
|---|---|---|---|
-h |
--help |
show this help message and exit | |
--dataset |
Cora |
Dataset to perform training on. Available options: Cora,CiteSeer,PubMedDiabetes | |
--emb-size |
128 |
Embedding dimension. Defaults to 128. | |
--reg-weight |
0.0 |
Weight to use for L2 regularization. If norm_reg is True, then reg_weight/num_of_nodes is used instead. | |
--norm-reg |
Boolean for whether to normalize the L2 regularization weight by the number of nodes in the graph. Defaults to false. | ||
--method |
node2vec |
Algorithm to perform training on. Available options: node2vec,GraphSAGE,GCN,DGI | |
--verbose |
1 |
Level of verbosity. Defaults to 1. | |
--epochs |
5 |
Number of epochs through the dataset to be used for training. | |
--optimizer |
Adam |
Optimization algorithm to use for training. | |
--learning-rate |
0.001 |
Learning rate to use for optimization. | |
--batch-size |
64 |
Batch size used for training. | |
--test-split |
0.1 |
Split of edge/non-edge set to be used for testing. | |
--output-fname |
None |
If not None, saves the hyperparameters and testing results to a .json file with filename given by the argument. | |
--node2vec-p |
1.0 |
Hyperparameter governing probability of returning to source node. | |
--node2vec-q |
1.0 |
Hyperparameter governing probability of moving to a node away from the source node. | |
--node2vec-walk-number |
50 |
Number of walks used to generate a sample for node2vec. | |
--node2vec-walk-length |
5 |
Walk length to use for node2vec. | |
--gcn-activation |
['relu'] |
Specifies layers in terms of their output activation (either relu or linear), with the number of arguments determining the length of the GCN. Defaults to a single layer with relu activation. | |
--graphSAGE-aggregator |
mean |
Specifies the aggreagtion rule used in GraphSAGE. Defaults to mean pooling. | |
--graphSAGE-nbhd-sizes |
[10, 5] |
Specify multiple neighbourhood sizes for sampling in GraphSAGE. Defaults to [25, 10]. |