6076paras/spd4gnns

SPD4GNNs

Implementation for Modeling Graphs Beyond Hyperbolic: Graph Neural Networks in Symmetric Positive Definite Matrices

Available Features

Graph Neural Networks

• GCNConv: Graph Convolutional Network

• GATConv: Graph Attentional Network

• SGCConv: Simplified Graph Convolutional Network

• GINConv: Graph Isomorphism Network

• ChebConv: Chebyshev-based Graph Convolutional Network

Manifolds

• SPD: The space of symmetric positive definite matrices - a special class of symmetric spaces.

• Euclidean: The space of vectors over the real field.

Datasets

• Disease: Disease Propagation Tree

• Airport: Flight Network

• Pubmed: Citation Network

• Citeseer: Citation Network

• Cora: Citation Network

Usage

Below are the instructions on how to run experiments on SPD and Euclidean spaces respectively.

python runner.py --dataset cora --model spdgcn --manifold spd --classifier linear

python runner.py --dataset cora --model gcn --manifold euclidean --classifier linear

Please note that the choice of datasets, models and manifolds remains open. Once chosen, the pipeline will load the optimal configuration of hyperparameters for the current setup in the 6-dimensional space, which can be found in the json folder (identified via grid-search).

Requirements

• Python == 3.8
• scikit-learn == 1.0.1
• torch == 1.12.1
• torch-geometric == 2.1.0

TODO

• Hyperbolic Space

• Datasets for Graph Classification