gravins / Anti-SymmetricDGN

Official code repository for the papers "Anti-Symmetric DGN: a stable architecture for Deep Graph Networks" accepted at ICLR 2023; "Non-Dissipative Propagation by Anti-Symmetric Deep Graph Networks"; and "Non-Dissipative Propagation by Randomized Anti-Symmetric Deep Graph Networks"

Anti-Symmetric DGN

This repository provides the official reference implementation of our papers

• Anti-Symmetric DGN: a stable architecture for Deep Graph Networks accepted at ICLR 2023
• Non-Dissipative Propagation by Anti-Symmetric Deep Graph Networks accepted at the DLG-AAAI’23 workshop.
• Non-Dissipative Propagation by Randomized Anti-Symmetric Deep Graph Networks accepted at the Deep Learning meets Neuromorphic Hardware Workshop, ECML-PKDD 2023.

Please consider citing us

@inproceedings{gravina2023adgn,
	author = {Alessio Gravina and Davide Bacciu and Claudio Gallicchio},
 	title = {{Anti-Symmetric DGN: a stable architecture for Deep Graph Networks}},
 	booktitle = {The Eleventh International Conference on Learning Representations },
 	year = {2023},
	url = {https://openreview.net/forum?id=J3Y7cgZOOS}
}

@inproceedings{gravina2023randomized,
	author = {Alessio Gravina and Claudio Gallicchio and Davide Bacciu},
 	title = {{Non-Dissipative Propagation by Randomized Anti-Symmetric Deep Graph Networks}},
 	booktitle = {},
 	year = {2023},
	url = {}
}

Requirements

Note: we assume Miniconda/Anaconda is installed, otherwise see this link for correct installation. The proper Python version is installed during the first step of the following procedure.

1. Install the required packages and create the environment

   • conda env create -f env.yml

2. Activate the environment

   • conda activate adgn

Run the experiment

Note: To run the experiment is fundamental to define the model name and dataset name into the run_all.sh files. For more details launch python3 main.py --help or python3 run-adgn-2-8.py --help

• Graph property prediction

  graph_prop_pred/run_all.sh

• Graph benchmarks

  graph_benchmark/run_all.sh

  • Note: please ensure that conf_seeds_5.json and conf_seeds_5.json match the generated seed before running the experiment

• Graph heterophilic benchmarks

  graph_heteropily/run_all.sh

• Tree-NeighborsMatch

  Tree-NeighborsMatch/run_all.sh

Repository structure

The repository is structured as follows:

├── README.md                <- The top-level README.
│
├── env.yml                  <- The conda environment requirements.
│
├── conda_list_output.txt    <- The specific version of each package in the adgn environment.
│
├── graph_benchmark          <- Contains the code to reproduce the graph benchmarks experiment.
│    ├── conf_seeds_5.json   <- The configuration seeds used in the experiment.
│    ├── data_seeds_5.json   <- The seeds used in the experiment to split the dataset into train/valid/test.
│    ├── run_all.sh          <- The script used to run the experiment. Note: you need to specify the name of the model and the dataset (see conf.py and utils/__init__.py)
│    ├── models              <- Contains the code for the framework A-DGN and other DGNs 
│    ├── utils               <- Contains the code for data and io utilities.
│    ├── main.py             <- The main.
│    ├── train.py            <- Implements the code responsible for training and evaluation of the models.
│    ├── conf.py             <- Contains the hyper-parameter space for each model.
│    └── model_selction.py   <- Implements the model selection and produces the report of the results.
│
└── graph_prop_pred          <- Contains the code to reproduce the graph property prediction experiment.
│   ├── run_all.sh           <- The script used to run the experiment. Note: you need to specify the name of the model and the dataset (see conf.py and utils/__init__.py)
│   ├── models               <- Contains the code for the framework A-DGN and other DGNs 
│   ├── utils                <- Contains the code for data and io utilities.
│   ├── main.py              <- The main.
│   ├── train_GraphProp.py   <- Implements the code responsible for training and evaluation of the models.
│   ├── conf.py              <- Contains the hyper-parameter space for each model.
│   └── model_selction.py    <- Implements the model selection and produces the report of the results.
│    
├── graph_heteropily          <- Contains the code to reproduce the graph heterophilic benchmarks experiment.
│    ├── run_all.sh          <- The script used to run the experiment. Note: you need to specify the name of the model and the dataset (see conf.py and utils/__init__.py)
│    ├── models              <- Contains the code for the framework A-DGN and other DGNs 
│    ├── utils               <- Contains the code for data and io utilities.
│    ├── main.py             <- The main.
│    ├── train.py            <- Implements the code responsible for training and evaluation of the models.
│    ├── conf.py             <- Contains the hyper-parameter space for each model.
│    └── model_selction.py   <- Implements the model selection and produces the report of the results.
│
└── Tree-NeighborsMatch      <- Contains the code to reproduce the Tree-NeighborsMatch experiment retrieved from the original source code: https://github.com/tech-srl/bottleneck.
    ├── run_all.sh           <- The script used to run the experiment. Note: you need to specify the name of the model and the dataset (see conf.py and utils/__init__.py)
    └── run-adgn-2-8.py      <- The main.