k-Redundancy Graph Neural Networks

On the Two Sides of Redundancy in Graph Neural Networks

This repository contains the code for paper On the Two Sides of Redundancy in Graph Neural Networks.

Usage

The following code demonstrates how to use the ToCanonizedDirectedAcyclicGraph class from the dag_gnn module to transform graph data into a DAG format. This transformation is useful in graph-related tasks such as node classification or graph classification.

from src.dag_gnn import ToCanonizedDirectedAcyclicGraph

k_redundancy = 0  # The hyperparameter that controls the redundancy of the graph
number_of_nodes = ...  # Set this to `None` for graph classification tasks
number_of_layers_for_your_dag_mlp_model = ...  # Must be greater than or equal to 1

dag_transform = ToCanonizedDirectedAcyclicGraph(num_nodes=number_of_nodes,
                                                num_layers=number_of_layers_for_your_dag_mlp_model,
                                                k=k_redundancy)
data = dag_transform(data)

# In the forward function, you extract the edges per layer:
def forward(self, data):
    dag_edge_index = data.dag_edge_index
    dag_layers_mask = data.dag_layers_mask
    dag_edge_attr = data.edge_multiplicities
    dag_readouts = data.dag_readout_at_each_layer
    dag_leaves_at_each_layer = data.dag_leaves_at_each_layer
    dag_x = data.dag_x
    leaves_0 = dag_leaves_at_each_layer[0]

    feature = ...  # Transform original node features `dag_x` to an embedding space

    x = zeros_like(feature)
    x[leaves_0] = feature[leaves_0]

    for i in range(self.num_layers):
        edge_index_i = dag_edge_index[:, dag_layers_mask == i]
        dag_edge_attr_i = dag_edge_attr[dag_layers_mask == i]
        # Additional operations per layer can be performed here
        ...

    x_at_each_layer = []
    for i in range(self.num_layers + 1):
        readout_i = dag_readouts[i]
        x_at_each_layer.append(x[readout_i])

Terms and conditions

When using our code please cite our paper:

@inproceedings{bause2024redundancy,
  title={On the Two Sides of Redundancy in Graph Neural Networks},
  author={Franka Bause, Samir Moustafa, Johannes Langguth, Wilfried N. Gansterer, and Nils M. Kriege},
  booktitle={ECML/PKDD, Lecture Notes in Computer Science 14946},
  year={2024},
}

Installation

Docker installation:

# 1. Build the docker image:
docker build -t kredundancygnn .
# 2. Run the docker container and attach to bash:
docker run -it --rm --gpus all kredundancygnn /bin/sh

Local installation:

# 1. (Optional) Create new environment for the the project:
conda create -n k_redundancy_gnn python=3.11     # it is needed to install python 3.10
# 2. (Optional) Activate the new environment:
conda activate k_redundancy_gnn
# 3. (Optional) Install cudatoolkit 11.3 and PyTorch dependencies (if you have GPU, otherwise skip this step):
conda install pytorch cudatoolkit=11.3 -c pytorch
# 4. Clone the repository:
git clone https://github.com/SamirMoustafa/k-RedundancyGNNs.git
# 5. Install the dependencies:
cd k-RedundancyGNNs && pip install -r requirements.txt
# 6. Export the repository path to PYTHONPATH *temporarily*:
export PYTHONPATH=$PYTHONPATH:$(pwd)

Reproducing the results

The results can be reproduced by running the following commands:

# Synthetic datasets (CSL, EXP)
python task_graph/synthetic/synthetic_case_study_run.py
python task_graph/synthetic/synthetic_run.py --dataset CSL
python task_graph/synthetic/synthetic_run.py --dataset EXP
# Node Classification (Cora, Citeseer, Pubmed, Cornell, Texas, Wisconsin)
python task_node/main_dagmlp.py
# TU datasets (IMDB-BINARY, IMDB-MULTI, ENZYMES, PROTEINS)
python task_graph/tudataset/tu_datasets_run.py

Repository Structure

.
|-- data
|   |-- CSL (with subdirectories for the data)
|   `-- EXP (with subdirectories for the data)
|-- src
|   |-- __init__.py
|   |-- canonized_dag.py
|   |-- canonized_ntrees.py
|   |-- dag.py
|   |-- dag_gnn.py
|   |-- hash_function.py
|   |-- ntrees.py
|   |-- process_daemon.py
|   `-- tensor_dag.py
|-- task_graph
|   |-- synthetic
|   |   |-- dagmlp_model.py
|   |   |-- dataset
|   |   |   |-- CSL.py
|   |   |   `-- EXP.py
|   |   |-- loader.py
|   |   |-- synthetic_case_study_run.py
|   |   `-- synthetic_run.py
|   `-- tudataset
|       |-- dagmlp_model.py
|       |-- statistics.py
|       `-- tu_datasets_run.py
|-- task_node
|   |-- dagmlp_model.py
|   |-- main_dagmlp.py
|   `-- main_gin.py
|-- test
|   |-- test_canonized_dag.py
|   |-- test_dag.py
|   |-- test_dag_mlp_layer.py
|   `-- test_hash_function.py
|-- requirements.txt
|-- Dockerfile
|-- README.md
`-- utils.py

Contact information

If you have any questions, please contact Franka Bause.