Neighbor2Seq: Deep Learning on Massive Graphs by Transforming Neighbors to Sequences

This repository is an official PyTorch implementation of Neighbor2Seq.

Meng Liu and Shuiwang Ji. Neighbor2Seq: Deep Learning on Massive Graphs by Transforming Neighbors to Sequences [SDM2022].

Requirements

• PyTorch
• PyTorch Geometric (with 1.6.1-1.7.2 recommended)
• OGB

Reference

@inproceedings{liu2022neighbor2seq,
  title={{Neighbor2Seq}: Deep Learning on Massive Graphs by Transforming Neighbors to Sequences},
  author={Liu, Meng and Ji, Shuiwang},
  booktitle={Proceedings of the 2022 SIAM International Conference on Data Mining},
  year={2022},
  organization={SIAM}
}

Run

All of our running scripts are included in run_ours.sh. An example on Flickr is as follows.

• Step 1: Precompute Neighbor2Seq

python precompute.py --dataset=Flickr --P=10 --add_self_loop=True

• Step 2: Train and evaluate Neighbor2Seq+Conv or Neighbor2Seq+Attn

CUDA_VISIBLE_DEVICES=0 python main_inductive.py --model=conv --lr=0.0008 --K=10 --weight_decay=0.00005 --hidden=256 --dropout=0.5 --batch_size=24576 --epochs=400 --kernel_size=7 --runs=10 --log_step=1 

CUDA_VISIBLE_DEVICES=0 python main_inductive.py --model=posattn --lr=0.002 --K=10 --weight_decay=0.00005 --hidden=256 --dropout=0.5 --batch_size=256 --epochs=200 --pe_drop=0.25 --runs=10 --log_step=1

Results

• Results on inductive tasks: Reddit, Flickr, and Yelp

• Results on ogbn-papers100M

• Results on ogbn-products