Graph Random Neural Features (GRNF) is an embedding method from graph-structured data to real vectors based on a family of graph neural networks. GRNF can be used within traditional processing methods or as a training-free input layer of a graph neural network. The theoretical guarantees that accompany GRNF ensure that the considered graph distance is metric, hence allowing to distinguish any pair of non-isomorphic graphs, and that GRNF approximately preserves its metric structure.

The following is the reference paper:

@inproceedings{zambon2020graph,
  title={Graph Random Neural Features for Distance-Preserving Graph Representations},
  author={Zambon, Daniele and Alippi, Cesare and Livi, Lorenzo},
  booktitle={Proceedings of the 37th International Conference on Machine Learning (ICML)},
  pages={10968--10977},
  editor={Hal Daumé III and Aarti Singh},
  volume={119},
  series={Proceedings of Machine Learning Research},
  address={Virtual},
  year={2020},
  month={13--18 Jul},
  publisher={PMLR},
  pdf={http://proceedings.mlr.press/v119/zambon20a/zambon20a.pdf},
}

GRNF has both a Spektral (TensorFlow) implementation:

from tensorflow.keras.layers import Input, Dense
from tensorflow.keras.models import Model
from grnf.tf import GraphRandomNeuralFeatures
X_in = Input(shape=(N, F))
A_in = Input(shape=(N, N))
psi = GraphRandomNeuralFeatures(64, activation="relu")([X_in, A_in])
output = Dense(1)(psi)
model = Model(inputs=[X_in, A_in], outputs=output)

as well as a PyTorch Geometric one:

from grnf.torch import GraphRandomNeuralFeatures
grnf = GraphRandomNeuralFeatures(64)
z = grnf(data)

Adapted from Spektral examples:

• examples/tf/delaunay_batch.py: Graph classification
• examples/tf/qm9_batch.py: Graph regression
• examples/tf/qm9_disjoint.py: Graph regression

Moreover, in examples/torch you can find also the code to test GRNF in the gnn-comparison framework.

Please, refer to branch v0.1.0

git checkout v0.1.0

• Improved pytorch implementation
• Added Spektral implementation working in batch and disjoint modes.
• Added examples

• Experiments of the paper