OGBL-Vessel Submission

This repo is the code on ogbl-vessel dataset.

Performance on ogbl-vessel (10 runs):

Methods	Test AUC	Valid AUC	Params
SAGE+JKNet (2-layers)	0.5001 ± 0.0033	0.5013 ± 0.0035	273
SAGE+JKNet (3-layers)	0.5001 ± 0.0007	0.5014 ± 0.0004	481
SAGE+JKNet (4-layers)	0.5003 ± 0.0005	0.5009 ± 0.0003	689
SEAL (no-xfeat)	0.8073 ± 0.0001	0.8077 ± 0.0001	43714

1. Setup

1.1 Environment

• Dependencies:

python==3.8
torch==1.10.1+cu102
torch-geometric==2.0.4
ogb==1.3.4
tqdm

• GPU: Tesla V100 (32GB)

1.2 Dataset

The dataset ogbl-vessel can be download and placed in ./dataset/ogbl_vessel/.

2. Usage

2.1 SAGE+JKNet

Full batch GraphSAGE that aggregates the outputs of each layer with JKNet.

• add JKNet (max): 3-layers and 4-layers make the performance more stable (std is lower) compared to raw GraphSAGE.
• reduce the dimension of hidden channels.
• set seed for implement.
• attempt leanrnable embeddings to replace data.x, but the performance is low.

To run SAGE+JKNet in 10 runs with seed 0-9:

cd scripts/
# 3 layers
bash train_sage.sh
# 4 layers
bash train_sage_4_hop.sh

2.2 SEAL

To run SEAL without using data.x in 10 runs with seed 0-9:

cd scripts/
bash train_seal.sh

3. Note

The implemention is based on https://github.com/snap-stanford/ogb/tree/master/examples/linkproppred/vessel and https://github.com/facebookresearch/SEAL_OGB.