This repository provides PyTorch implementations for graph classification using Graph Convolutional Networks (GCN), Graph U-Net, and Multigraph GCN (MGCN). It aims to reproduce and compare results from recent research papers, offering a baseline for evaluating graph neural network architectures on datasets like PROTEINS and ENZYMES.

How It Works

The core approach involves implementing graph convolutional layers and pooling mechanisms for graph classification. Graph U-Net utilizes a pooling strategy based on node dropping between graph convolution layers, differing from standard GCNs by simplifying the readout layer to max pooling. The implementation is basic, prioritizing clarity for debugging and experimentation over extensive optimization.

Quick Start & Requirements

• Install via pip (PyTorch 0.4.1/1.0.0, Python 3.6 tested on Ubuntu 16.04).
• Optional: PyTorch Geometric data loader via -g flag.
• Datasets are included or can be downloaded.
• See README for example commands and results parsing.

Highlighted Details

• Compares GCN, Graph U-Net, and MGCN performance on PROTEINS and ENZYMES datasets.
• Includes results with and without continuous node attributes.
• Supports one-hot node degree features via --degree flag.
• Offers basic implementation for debugging and experimentation.

Maintenance & Community

• Project by Boris Knyazev.
• Last updated in 2018.
• References research papers from 2017-2018.

Licensing & Compatibility

• No explicit license mentioned in the README.
• Code is presented as an attempt to reproduce research results.

Limitations & Caveats

The decoder part of Graph U-Net is not implemented. Performance may be affected by the basic implementation without optimizations. The project appears to be a research artifact from 2018, with no indication of ongoing maintenance or updates.