vGraph: A Generative Model For Joint Community Detection and Node Representational Learning

This is a Pytorch implementation of the paper vGraph: A Generative Model For Joint Community Detection and Node Representational Learning and is done under the NeurIPS Reproducibility Challenge 2019. The original implementation by author can be found here.

Summary of the paper

This paper proposes a novel technique for learning node representations and at the same time perform community detection task for the graphical data by creating a generative model using the variational inference concepts. The full paper summary along with its main contributions can be found here

Setup Instructions and Dependancies

The code has been written in Python 3.6 and Pytorch v1.1. Also Pytorch Geometric has been used for training procedures, along with the usage of TensorboardX for logging loss curves.

For training/testing the model, you must first download Facebook social circles dataset. It can be found here. After downloading the dataset, all the files must be placed inside ./dataset/Facebook/.

Repository Overview

The following is the information regarding the various important files in the directory and their functions:

• model.py: File containing the network architecture
• utils.py: File containing helper functions and losses
• data.py: File containing functions to call dataset in an operable format
• train_nonoverlapping.py: File containing the training procedure for non-overlapping dataset
• train_overlapping.py: File containing the training procedure for overlapping dataset

Running the model

For training the model, use the following commands:

python train_nonoverlapping.py            ### For training non-overlapping dataset
python train_overlapping.py               ### For training overlapping dataset

Current Status of the Project

Currently the directory contains dataloader and training procedure for 2 non-overlapping datasets(Cora and Citeseer) and 10 overlapping datasets(facebook0, facebook107, facebook1684, facebook1912, facebook3437, facebook348, facebook3980, facebook414, facebook686, facebook698). I plan to add more dataloaders in the directory. Also the various accuracy measures as specified in the paper will also soon be added in the repository.

If you found the codebase useful in your research work, consider citing the original paper

License

This repository is licensed under MIT License