This repository provides evaluation codes of CFG on ogbl-citation2 dataset for OGB link property prediction task. The idea of CFG is described in the following article:
Circle Feature Graphormer: Can Circle Features Stimulate Graph Transformer?
This implementation of CFG for Open Graph Benchmak datasets (ogbl-citation2) is based on OGB and SIEG. Thanks for their contributions.
The code is implemented with PyTorch and PyTorch Geometric.
Requirments:
- python=3.7.11
- pytorch=1.10.0
- ogb=1.3.6
- torch-geometric=1.7.2
- dgl=0.8.1
Install PyTorch
Install PyTorch_Geometric
Install Networkx
Install OGB
Install DGL
Other required python libraries include: numpy, scipy, tqdm etc.
python3 train.py --ngnn_code --grpe_cross --device 0 --num_heads 8 --dataset ogbl-citation2 --use_feature --use_feature_GT --use_edge_weight --epochs 20 --train_percent 8 --val_percent 4 --test_percent 0.2 --model NGNNDGCNNGraphormer_noNeigFeat --runs 10 --batch_size 64 --lr 2e-05 --num_workers 24 --dynamic_train --dynamic_val --dynamic_test --use_len_spd --use_num_spd --use_cnb_jac --use_cnb_aa --use_cnb_swing
or
sh train_citation2.sh
python3 train.py --ngnn_code --grpe_cross --device 0 --num_heads 8 --dataset ogbl-citation2 --use_feature --use_feature_GT --use_edge_weight --epochs 20 --train_percent 8 --val_percent 4 --test_percent 0.2 --model NGNNDGCNNGraphormer_noNeigFeat --runs 10 --batch_size 64 --lr 2e-05 --num_workers 24 --dynamic_train --dynamic_val --dynamic_test --use_len_spd --use_num_spd --use_cnb_jac --use_cnb_aa --use_cnb_bridge
The performances of CFG together with some selected GNN-based methods on OGB-CITATION2 task are listed as below:
| Method | Test MRR | Validation MRR |
|---|---|---|
| PLNLP | 0.8492 ± 0.0029 | 0.8490 ± 0.0031 |
| AGDN w/GraphSAINT | 0.8549 ± 0.0029 | 0.8556 ± 0.0033 |
| SEAL | 0.8767 ± 0.0032 | 0.8757 ± 0.0031 |
| S3GRL (PoS Plus) | 0.8814 ± 0.0008 | 0.8809 ± 0.0074 |
| SUREL | 0.8883 ± 0.0018 | 0.8891 ± 0.0021 |
| NGNN + SEAL | 0.8891 ± 0.0022 | 0.8879 ± 0.0022 |
| SIEG | 0.8987 ± 0.0018 | 0.8978 ± 0.0018 |
| CFG1 | 0.8997 ± 0.0015 | 0.8987 ± 0.0011 |
| CFG2 | 0.9003 ± 0.0007 | 0.8992 ± 0.0007 |
CFG achieves top-1 performance on ogbl-citation2 in current OGB Link Property Prediction Leader Board until Sep 14, 2023.
CFG is released under an MIT license. Find out more about it here.