radoslav11/SP-MPNN

Implementation of the SPN model and the experiments from the LoG 2022 paper "Shortest Path Networks for Graph Property Prediction".

Shortest Path Message Passing Neural Networks

This repository contains the source code for the Shortest Path Message Passing Neural Network (SP-MPNN) framework models, presented in the LoG 2022 paper "Shortest Path Networks for Graph Property Prediction". The repository includes the h-Prox datasets, as well as all evaluation datasets and code for training and testing the different SP-MPNN variants across all experiments in the paper.

Requirements

The requirements for the Python environment can be found in requirements.txt. The main packages that have been used are PyTorch, PyTorch Geometric, OGB (Open Graph Benchmark), Neptune (neptune.ai) and Numpy.

Running on a GPU is supported via CUDA.

Datasets

TU Datasets (Chemical)

Available in Pytorch Geometric.

Proximity

The synthetic Prox datasets (k = {1, 3, 5, 8, 10}) are available here. To use these datasets, the contents of the aforementioned zip file must be extracted and moved into the data directory

MoleculeNet

We use the OGB implementation of the MoleculeNet datasets. More information can be found here.

QM9

The QM9 dataset is provided in data/QM9.

Running

The script we use to run the experiments is src/main.py. Note that the script should be run from inside the src directory, or mark it as Source Root.

The main parameters of the script are:

• --dataset the dataset we use.
• --model for the model used. The main convention is SP-{INNER}-{OUTER}, where INNER corresponds to the approach we use for aggregating the embeddings on each hop level, while OUTER is the approach we use for aggregating the different hop levels. The main models we use are SP-SUM-WEIGHT, SP-EDGESUM-WEIGHT and SP-RSUM-WEIGHT, where the common WEIGHT outer aggregation is the normalised sum that the simple model in the paper uses (SPN).
• --mode for the current task type. We use gc for Graph Classification, and gr for Graph Regression.

Additionally, some of the more useful configurable parameters are:

• --emb_dim for the embedding dimensionality.
• --batch_size for the batch size we use during training.
• --lr for the learning rate.
• --dropout for the dropout probability.
• --epochs for the number of epochs.

A detailed list of all additional arguments can be seen using the following command:

python main.py -h

Example running

You can run the end-to-end experiments for the datasets from each class using the commands below, where the arguments are replaced with:

• {dataset} is the dataset we want to run the experiment on
• {k} is the maximum distance for a SP layer
• {L} is the number of layers
• {QM9_TASK} is an integer id for the QM9 objective / task we want to predict

TU Datasets (Chemical), Prox

python main.py -d {dataset} -m SP-SUM-WEIGHT --max_distance {k} --num_layers {L} --mode gc

MoleculeNet

python main.py -d {dataset} -m SP-EDGESUM-WEIGHT --max_distance {k} --num_layers {L} --mode gc

QM9

python main.py -d QM9 -m SP-RSUM-WEIGHT --max_distance {k} --num_layers {L} --specific_task {QM9_TASK} --mode gr

Neptune.ai

You can use neptune.ai to track the progress, by specifying your project and token in src/config.ini. Leave the fields as ... if you want to just run locally.

Citing this paper

If you make use of this code, or its accompanying paper, please cite this work as follows:

@inproceedings{ADC-LoG2022,
  title={Shortest Path Networks for Graph Property Prediction},
  author    = {Ralph Abboud and Radoslav Dimitrov and
                {\.I}smail {\.I}lkan Ceylan},
  booktitle={Proceedings of the First Learning on Graphs Conference ({LoG})},
  year={2022},
  note={Oral presentation}
}