Training-free Graph Neural Networks and the Power of Labels as Features (TMLR 2024)

We propose training-free graph neural networks (TFGNNs), which can be used without training and can also be improved with optional training.

Paper: https://arxiv.org/abs/2404.19288

💿 Preparation

Install Poetry and run the following command:

$ poetry install

🧪 Evaluation

TFGNNs Outperform Existing GNNs in Training-free Setting

$ poetry run python trainingfree.py

Model	Cora	CiteSeer	PubMed	CS	Physics	Computers	Photo
GCNs	0.163	0.167	0.180	0.079	0.101	0.023	0.119
GCNs + LaF	0.119	0.159	0.407	0.080	0.146	0.061	0.142
GATs	0.177	0.229	0.180	0.040	0.163	0.058	0.122
GATs + LaF	0.319	0.077	0.180	0.076	0.079	0.025	0.044
TFGNNs + random initialization	0.149	0.177	0.180	0.023	0.166	0.158	0.090
TFGNNs (proposed)	0.600	0.362	0.413	0.601	0.717	0.730	0.637

▲ Results. TFGNNs outperform existing GNNs in the training-free setting.

Deep TFGNNs Perform Better in Training-free Setting

$ poetry run python numlayers.py

▲ Results. Deep TFGNNs perform better in the training-free setting.

TFGNNs Converge Fast in Optional Training

$ poetry run python training.py

▲ Results. TFGNNs converge fast in optional training.

TFGNNs are Robust to Feature Noise

$ poetry run python noisy.py

▲ Results. TFGNNs are robust to feature noise.

🖋️ Citation

@article{sato2024trainingfree,
  author    = {Ryoma Sato},
  title     = {Training-free Graph Neural Networks and the Power of Labels as Features},
  journal   = {Transactions on Machine Learning Research},
  year      = {2024},
}