VP-GNN

This is the code for paper Augmenting Message Passing with Virtual Paths via Effective Resistance.

TL;DR

• VP-GNN is a novel graph learning framework that effectively alleviates over-squashing without impairing the inherent graph structure.

Main Requirements

• Python 3.10.13

  • conda create -n your_env_name python=3.10.13

• torch 2.1.1 + cu118

  • pip install torch==2.1.1 torchvision==0.16.1 torchaudio==2.1.1 --index-url https://download.pytorch.org/whl/cu118

• Install causal_conv1d and mamba (you should download this from Vim)

  • pip install -e causal_conv1d>=1.1.0
  • pip install -e mamba-1p1p1

• Install pyG:

  • pip install torch_geometric
  • pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu118.html

• Install dgl: -pip install dgl==0.9.0

How to Run

Node Classification

Example Command:

python run_node_classification.py --rewiring=VP --dataset={$dataset} --layer_type={$model} --num_path={$num_path} --device={$device}

The arguments we consider in our paper:

• dataset:
  • cora, texas, citeseer, wisconsin, cornell, squirrel
• model:
  • VPaugGCN, VPaugGIN
• num_path:
  • [5, 10, 15, 20, 25, 30]
• device:
  • gpu number

If you want to run 100 tirals, add num_trials like below:

python run_node_classification.py --rewiring=VP --dataset={$dataset} --layer_type={$model} --num_path={$num_path} --device={$device}--num_trials=100