Variational Graph Auto-Encoders

The VGAE-based model is an unsupervised-based model which takes as enconder the Graph Convolutional Network (GCN) model. VGAE extracts latent variables from a given connected graph, then the decoding process is realized through a simple inner product.

1. Files

train.py: Main file to train the model.
src/model.py: Contains the VGAE class.
src/utils.py: Contains functions to load raw files and transform them into a graph format.
src/parser.py: Contains parser values.
benchmarks/: Holds datasets as csv format.

1.1 Datasets

All datasets will be stored in the directory benchmarks/. Inside benchmark/ you need to store the directory of your dataset such as benchmarks/[your_dataset_directory]/, inside of [your_dataset_directory/] you have to store the graph file with csv format such as: dataset.csv.

The dataset.csv needs to meet the following format:

  node1 node2
  node2 node3
  node3 node4
  ...

2. Installation

It is recommended to use a virtual environment such as pipenv to install the dependencies and run the model. If you do not have installed pipenv yet, just type: pip install pipenv. Then, you need to launch the virtual environment by typing:

pipenv shell

Once the virtual environment has been launched, you need to install the dependencies:

pipenv install

in case the previous command does not work, try to type the next command:

pipenv install --ignore-pipfile

3. How to run

To run the model you need to use the file train.py such as:

python train.py [--embedding_size EMBEDDING_SIZE] [--epochs EPOCHS]
                [--learning_rate LEARNING_RATE] [--neurons NUM_NEURONS]
                [--dataset DATASET] [--directory DIRECTORY]
                [--test_size TEST_SIZE]

4. Further work

The model can be improved by adding:

A function to visualize embeddings
A function to evaluate results

Note: Feel free to send me a pull request if you want to collaborate in this proyect.