Intro

The code is for the paper Anonymous Walk Embeddings (https://arxiv.org/abs/1805.11921) that creates embeddings for entire graphs, based on feature-based and data-driven approaches, and evaluates the results on graph classification task.

Tutorial

Check a notebook Tutorial.ipynb to see how to get network embeddings, load/save embeddings matrix, calculate kernel matrix, and perform SVM calculation.

Running code

The following command will run data-driven model with the parameters.

python AnonymousWalkEmbeddings.py --dataset mutag --root ./Datasets --window_size 16 --batch_size 100 --batches_per_epoch 100 --num_samples 64 --steps 10 --results_folder mutag_results

The following command will run data-driven model with the parameters.

python AnonymousWalkKernel.py --dataset mutag --root ./Datasets --steps 10 --method sampling --MC 10000 --results_folder mutag_results

Alternatively, you can change default parameters in the file doc2vec.py.

Below is the list of all arguments:

• dataset Name of the dataset (and folders to save/load data)
• batch_size Number of (context, target) pairs per batch
• window_size Number of words in the context
• embedding_size_w Dimension of word embeddings
• embedding_size_d Dimensions of doc embeddings
• num_samples number of (negative) samples for every target word.
• concat Concatenate or Average embeddings of words.
• loss_type Sampled softmax loss or NCE loss
• optimize Adagrad or SGD
• learning_rate Learning rate
• root Location of dataset
• ext Extension of the graph files
• steps Number of steps in a random walk.
• epochs Number of global iterations.
• batches_per_epoch number of batches per epoch for each graph
• candidate_func None (loguniform by default) or uniform
• graph_labels None, edges, nodes, edges_nodes
• method Sampling or Exact.
• MC Number of Monte-Carlo iterations.
• results_folder Folder for storing results.

Running with Docker

To run the code on docker, we first need to obtain an appropriate image. After installing Docker on your computer, you first pull an image gcr.io/tensorflow/tensorflow for Tensorflow:

docker pull gcr.io/tensorflow/tensorflow

After that you need to build a Dockerfile. You should have in the current directory Dockerfile. To avoid compression of all the files under current directory, we recommend to create a separate folder for Dockerfile and navigate there.

cd docker
docker build -t awe-docker .

You can test available images on your computer with docker images. Once you have awe-docker image you can run the code with.

docker run -i --user $(id -u):$(id -g) --mount type=bind,source=/home/sivanov/Datasets,target=/src/Datasets --mount type=bind,source=/home/sivanov/awe/,target=/src/awe/ --name='sergey.mutag.1' -t awe-docker:latest python /src/awe/AnonymousWalkEmbeddings.py --dataset mutag --root /src/Datasets/ --window_size 16 --batch_size 100 --batches_per_epoch 100 --num_samples 64 --results_folder /src/awe/mutag_results

Here, --user $(id -u):$(id -g) makes sure your results will be written with your user permission. --mount type=bind,source=/home/sivanov/Datasets,target=/src/Datasets mounts your host directory with datasets into docker image under /src/Datasets directory. --mount type=bind,source=/home/sivanov/awe/,target=/src/awe/ mounts your host directory /home/sivanov/awe/ with scripts into an image directory /src/awe/. --name='sergey.mutag.1' sets the name of container. -t awe-docker:latest provides a name of the image.

After running, you should get results in your host machine at awe/mutag_results directory.

Citation

If you use the code, please consider citing our work.

@InProceedings{pmlr-v80-ivanov18a,
  title = 	 {Anonymous Walk Embeddings},
  author = 	 {Ivanov, Sergey and Burnaev, Evgeny},
  booktitle = 	 {Proceedings of the 35th International Conference on Machine Learning},
  pages = 	 {2191--2200},
  year = 	 {2018},
  editor = 	 {Dy, Jennifer and Krause, Andreas},
  volume = 	 {80},
  series = 	 {Proceedings of Machine Learning Research},
  address = 	 {Stockholmsmässan, Stockholm Sweden},
  month = 	 {10--15 Jul},
  publisher = 	 {PMLR},
  pdf = 	 {http://proceedings.mlr.press/v80/ivanov18a/ivanov18a.pdf},
  url = 	 {http://proceedings.mlr.press/v80/ivanov18a.html}
  }