adocherty/node2vec_experiments

node2vec_experiments

Computes and evaluates node2vec embeddings from context pairs file.

How to run the code:

The code takes a text file containing target-context pairs of node indices:

1 3
40 2
4 5
...

Currently, the filenames are hard coded in the file. It expects two files:

• gPairs-w3-s6.txt
• karate-labels.txt

To change this, edit the following lines:

 ds = PregeneratedDataset(<target context pairs filename>,
                          n_nodes=<number of nodes in graph>,
                          delimiter="\t",
                          force_offset=-1,
                          splits=[0.8,0.2])

Additionally changing the delimiter and offset to match the file. The force_offset flag is added to the node indices in the target-context file to map the file node indices to the range 0 to (n_nodes-1).

Training using previous node embeddings:

Node embeddings are automatically stored in tensorflow checkpoint files, they can be reloaded in the code by specifying the checkpoint_file. The checkpoints files are named checkpoint... for incremental checkpoints and model_epoch_<n> saved after epoch n.

e.g.

checkpoint_file = "n2v_2018-04-18/checkpoint-170"

Freezing node embeddings

Training can be performed with some embeddings frozen.

To freeze a set of node embeddings use:

freeze_indices = [199, 200, 399, 400]

To freeze a set of context embeddings (but allow embeddings to change) use:

freeze_context_indices = [199, 200, 399, 400]