Graph Transformer Structural Encoding Based on MCL Algorithm

We propose a general structural encoding method for graph Transformer.And experimented on the OGB graph classification task ogbg-ppa.

Requirements

ogb==1.3.3 torch==1.10.2 torch_geometric==2.0.4 tqdm==4.64.0

Training & Evaluation

ogbg-ppa

# $INFLA and $OUTPUT are described below.
python main.py --infla $INFLA --output $OUTPUT

• $INFLA: MCL encoding granularity can be changed by changing this parameter. Leaving this parameter unfilled means that the default INFLA(6) is used. This parameter can also be entered as an array of float types for a multi-scale encoding effect(For example: --infla 3.0 6.0 9.0).
• $OUTPUT: Specifying output files. The contents of the output file contain the hyperparameters used in the training, as well as the final Train ACC, Validation ACC and Test ACC.

Performance

ogbg-ppa

Methods	Train ACC	Validation ACC	Test ACC	Hardware
GC-T+MCL(6.0)	0.999719 $\pm$ 0.000318	0.698900 $\pm$ 0.003674	0.743190 $\pm$ 0.003319	RTX3090

References

1.https://github.com/snap-stanford/ogb/tree/master/examples/graphproppred/ppa