This repository contains an implementation of the Variational Graph Auto-Encoder (VGAE) as described in the paper by T. Kipf and M. Welling, "Variational Graph Auto-Encoders", published on arXiv in 2016. The VGAE is an approach for unsupervised learning on graph-structured data, demonstrating superior performance on the link prediction task.
The implementation focuses on the Cora dataset, adhering to the default settings specified in the original paper. This includes:
- Training for 200 epochs.
- Utilizing the
torch.optim.Adam()optimizer with a learning rate of 0.02. - A two-layer Graph Convolutional Network (GCN) for the encoder, with additional layers for computing
mu(the mean) andlog_std(the log standard deviation). - Hidden layer size of 32 and a latent space size of 16.
- Employing the reparametrization trick for latent space embedding generation.
- A simple inner product decoder for reconstructing the graph.
To set up the project for local development and testing, follow this:
git clone https://github.com/dorochka8/VGAE.git
To train the model with the default settings on the Cora dataset, run:
python3 main.py
You can modify the training parameters within the config.py script to experiment with different configurations.
The results are provided for 10 independent training loops, each running for 200 epochs. Below is a table showcasing the loss, ROC AUC (Receiver Operating Characteristic Area Under the Curve), and AP (Average Precision) for 10 independent training loops:
| Loss | ROC AUC | AP |
|---|---|---|
| 0.256 | 0.906 | 0.900 |
The mean ROC AUC across the training loops is 0.906, indicating a high true positive rate relative to the false positive rate. The mean Average Precision (AP), which measures the precision-recall tradeoff, is 0.9, demonstrating the model's effectiveness in link prediction within the graph.
These results underscore the VGAE model's robustness and efficiency in handling graph-structured data, closely aligning with the benchmarks set by the original paper. The consistent performance across multiple runs highlights the model's stability and reliability for unsupervised learning tasks on graph data.