This is a pytorch implementation of the paper Scalable Expressiveness through Preprocessed Graph Perturbations.
Please read our paper at: https://arxiv.org/abs/2406.11714v1
We introduce Scalable Expressiveness through Preprocessed Graph Perturbation (SE2P), a model combining flexibility, scalability, and expressiveness. Our approach offers four configuration classes, each offering a unique balance between scalability and generalizability. Through preprocessing, SE2P generates multiple perturbations of the input graph by a perturbation policy (e.g., random node removal) and diffuses nodal features across each perturbed graph. SE2P leverages the expressive power offered by multiple perturbed graphs to surpass 1-WL expressiveness limits. The flexibility of SE2P is of practical importance, allowing for the selection of learnable or non-learnable aggregation functions and thus enabling scalable or expressive variations of many models.
- Python 3.8
- For the other packages, please refer to the requirements.txt.
To run SE2P configuration on the TU benchmark datasets run SE2P.py. You can specify the following arguments:
dataset- The dataset from the TU benchmark, including MUTAG, PROTEINS, PTC_GIN, IMDB-BINARY, IMDB-MULTI, and COLLAB.configuration- Which configurations to select, options are c1, c2, c3, and c4.L- The number of (virtual) layers in the diffusion step.n_f- Number of layers in the final MLP (decoder).n_p- Number of layers in the MLP of the POOL function. Needed for c2, c3, and c4.ds_mi- Number of layers in the inner MLP of the DeepSet for the MERGE function. Needed for c3 and c4.ds_mo- Number of layers in the outer MLP of the DeepSet for the MERGE function. Needed for c3 and c4.ds_ci- Number of layers in the inner MLP of the DeepSet for the COMBINE function. Needed for c4.ds_co- Number of layers in the outer MLP of the DeepSet for the COMBINE function. Needed for c4.h- The hidden dimentionality of the hidden layers.graph_pooling- If set to "attention_agg", attentional aggregation will be used for the POOL function. In our experiments, "sum" is used.batch_size- The batch size of the model.seed- Seed for reproducibilty. Default is 0.epochs- Number of epochs for training. For TU datasets 350 is used.batch_norm- Whether to use batch normalization or not.
To run SE2P on the OGB benchmark datasets, run OGBG.py. All the arguments are same as TU benchmark, except for the following:
dataset- Select either "ogbg-molhiv" or "ogbg-moltox21".epochs- Number of epochs, which 100 is used on this benchmark.
The DROPGNN.py file, which aims at reproducing the results of the DropGNN paper, is based on "Dropout Graph Neural Network" code by Papp et al. (https://github.com/KarolisMart/DropGNN)
Please cite our work if you find it useful in any way.
@article{saber2024scalable,
title={Scalable Expressiveness through Preprocessed Graph Perturbations},
author={Saber, Danial and Salehi-Abari, Amirali},
journal={arXiv preprint arXiv:2406.11714},
year={2024}
}