dsgiitr / graph_nets

PyTorch Implementation and Explanation of Graph Representation Learning papers: DeepWalk, GCN, GraphSAGE, ChebNet & GAT.

Graph Representation Learning

This repo is a supplement to our blog series Explained: Graph Representation Learning. The following major papers and corresponding blogs have been covered as part of the series and we look to add blogs on a few other significant works in the field.

Setup

Clone the git repository :

git clone https://github.com/dsgiitr/graph_nets.git

Python 3 with Pytorch 1.3.0 are the primary requirements. The requirements.txt file contains a listing of other dependencies. To install all the requirements, run the following:

pip install -r requirements.txt

1. Understanding DeepWalk

Unsupervised online learning approach, inspired from word2vec in NLP, but, here the goal is to generate node embeddings.

• DeepWalk Blog
• Jupyter Notebook
• Code
• Paper -> DeepWalk: Online Learning of Social Representations

2. A Review : Graph Convolutional Networks (GCN)

GCNs draw on the idea of Convolution Neural Networks re-defining them for the non-euclidean data domain. They are convolutional, because filter parameters are typically shared over all locations in the graph unlike typical GNNs.

• GCN Blog
• Jupyter Notebook
• Code
• Paper -> Semi-Supervised Classification with Graph Convolutional Networks

3. Graph SAGE(SAmple and aggreGatE)

Previous approaches are transductive and don't naturally generalize to unseen nodes. GraphSAGE is an inductive framework leveraging node feature information to efficiently generate node embeddings.

• GraphSAGE Blog
• Jupyter Notebook
• Code
• Paper -> Inductive Representation Learning on Large Graphs

4. ChebNet: CNN on Graphs with Fast Localized Spectral Filtering

ChebNet is a formulation of CNNs in the context of spectral graph theory.

• ChebNet Blog
• Jupyter Notebook
• Code
• Paper -> Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering

5. Understanding Graph Attention Networks

GAT is able to attend over their neighborhoods’ features, implicitly specifying different weights to different nodes in a neighborhood, without requiring any kind of costly matrix operation or depending on knowing the graph structure upfront.

• GAT Blog
• Jupyter Notebook
• Code
• Paper -> Graph Attention Networks

Citation

Please use the following entry for citing the blog.

@misc{graph_nets,
  author = {A. Dagar and A. Pant and S. Gupta and S. Chandel},
  title = {graph_nets},
  year = {2020},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/dsgiitr/graph_nets}},
}

graph_nets's Issues

Is unsupervised learning implemented for graphSage?

Is unsupervised learning implemented for graphSage?

Problems about graphsage

In GraphSage part, just one hop neighbors were used in enc1 and enc2.
Maybe , using 2 hop neighbors in enc1 and 1 hop neighbors in enc2 is more proper.
Just a little suggestion.

ChebNet dataset

I saw the code has this code sentence

how can i get this datasets folder

Question: GCN pytorch implementation

Thanks a lot for the code. Found it via your blog post about GCNs.

I see Kipf used Laplace symmetric normalisation - as you wrote also here: https://dsgiitr.com/blogs/gcn/

But I don't understand, why the diagonal degree matrix is produced using A instead of A_hat:

        self.A_hat = A+torch.eye(A.size(0))
        self.D     = torch.diag(torch.sum(A,1))  #<< should this not be self.A_hat?
        self.D     = self.D.inverse().sqrt()