Graph-Neural-Network-

Tensorflow and Pytorch implementation of a Graph Neural Network (GNN) model, referred to as GNN+ in this documentation.

The GNN+ model has versatile applications, including vertex clustering and graph classification. It excels in creating precise cluster assignments and can be a valuable tool for various graph-related tasks.

Model Description

GNN+ comprises two essential components:

GNNConv

The GNNConv layer is a message-passing layer designed to minimize the $L_1$-norm difference between features of adjacent nodes. It updates node features based on the graph topology and hyperparameters, leading to improved cluster assignments.

CheegerCutPool

The CheegerCutPool is a graph pooling layer equipped with an internal MLP. It calculates cluster assignments and produces coarsened graphs for further processing.

Downstream Tasks

GNN+ is suitable for both unsupervised vertex clustering and supervised graph classification. Its flexibility allows it to tackle a wide range of tasks in graph analysis.

Vertex Clustering

In the unsupervised task of vertex clustering, GNN+ partitions vertices into clusters based on their features and graph topology. The model is trained using an unsupervised loss.

Graph Classification

Graph classification is a supervised task aiming to predict the class of each graph. GNN+ alternates GNNConv layers with graph pooling layers to extract global label information from vertex representations.

Implementation

Tensorflow

This implementation is based on the Spektral library and follows the Select-Reduce-Connect API. To execute the code, first install the conda environment from tf_environment.yml

conda env create -f tf_environment.yml

The tensorflow/ folder includes:

• The implementation of the GNNConv layer
• The implementation of the CheegerCutPool layer
• An example script to perform the clustering task
• An example script to perform the classification task

Pytorch

This implementation is based on the Pytorch Geometric library. To execute the code, first install the conda environment from pytorch_environment.yml

conda env create -f pytorch_environment.yml

The pytorch/ folder includes:

• The implementation of the GNNConv layer
• The implementation of the CheegerCutPool layer
• An example script to perform the clustering task
• An example script to perform the classification task

Spektral

GNN+ is now available on Spektral:

• GNNConv layer,
• CheegerCutPool layer,
• Example script to perform node clustering with GNN+.