Graph Convolution Network Classifier with Sparse Adjacency Matrix Prior

Please click on the ‘Code’ button displayed above to view project source code in GitHub.

Abstract

Benchmarking performances of Fully Connected Network (FCN), Deep Convolutional Neural Network (CNN) & Graph Convolution Network with 3 different priors on MNIST.

Available Models

1. Fully Connected
2. Graph (with my Custom Adj Matrix - default)
3. Convolution
4. Graph (with Gaussian Adj Matrix)
5. Graph (with Trainable Adj Matrix)

Implementation of a Graph Neural Network Classifier with 3 different priors

1. Sparse Adjacency Matrix (Feng et al., 2020)
2. Gaussian Adjacency Matrix & Normalization as per (Kipf & Welling et al., ICLR 2017)
3. Trainable Adjacency Matrix (Predict Edges)

Requirements

1. Python 3
2. PyTorch
3. Numpy
4. Networkx
5. Scipy

Usage

1. python graph_neural_network.py --model fc
2. python graph_neural_network.py --model conv
3. python graph_neural_network.py --model gaussian_graph
4. python graph_neural_network.py --model graph --pred_edge
5. python graph_neural_network.py --model graph