SSNP: Stochastic Subgraph Neighborhood Pooling

SSNP (Stochastic Subgraph Neighborhood Pooling) is a fork of GLASS. SSNP uses subgraph neighborhoods during pooling to increase the expressiveness of plain-GNNs on subgraph classification. SSNP jointly aggregates the subgraph and its neighborhood (i.e., external topology) information without any computationally expensive operations such as labeling tricks. To improve scalability and generalization further, SSNP also uses a simple data augmentation pre-processing step that creates multiple sparse views of the subgraph neighborhood.

arXiv link: https://arxiv.org/pdf/2304.08556.pdf

Prepare Data

The realworld datasets can be downloaded from here.

Commands to run our model

To run Online Views (OV):

python ssnp.py --use_nodeid --use_seed --repeat 10 --epochs $epochs --dataset $dataset --model 2 --samples 1 --m $m --M $M --stochastic --device $gpu_id

To run Pre-processed Views (PV):

python ssnp.py --use_nodeid --use_seed --repeat 10 --epochs $epochs --dataset $dataset --model 2 --samples 1 --m $m --M $M --views $views --device $gpu_id

To run Pre-processed Online Views (POV):

python ssnp_hybrid.py --use_nodeid --use_seed --repeat 10 --epochs $epochs --dataset $dataset --model 2 --m $m --M $M --nv $nv --nve $nve --device $gpu_id

Command line arguments

The following command line arguments are supported by our model

• dataset - Set the dataset (currently supported datasets are ppi-bp, hpo-metab, hpo-neuro and em-user)
• use_nodeid - Use pretrained node embeddings (found in Emb folder)
• repeat - Set number of times experiment should be run
• epochs - Set the number of epochs for the model
• m - Set the length of the random walk (called h in the paper)
• M - Set the number of random walks (called k in the paper)
• device - Set the GPU ID to be used (eg; 0)
• nv - Set the total number of views
• nve - Set the number of views per epoch (only for POV)
• diffusion - An additional MLP after the pooling layer to mix the subgraph and neighborhood information

Configs

The dataset configurations such as pooling function, learning rate, number of convolution layers etc., can be set in compl-config/$dataset.

Bibtex

If you find our work useful, please cite us using the following:

@article{jacob2023stochastic,
  title={Stochastic Subgraph Neighborhood Pooling for Subgraph Classification},
  author={Jacob, Shweta Ann and Louis, Paul and Salehi-Abari, Amirali},
  journal={arXiv preprint arXiv:2304.08556},
  year={2023}
}