SmilesDZgk / MGC

Code for MGC

This repo contains an example implementation of our paper (under reviewing for TNNLS) "Handling Over-smoothing and Over-squashing in Graph Convolution with Maximization Operation" . Please note that the repository is not yet finalized and will be further enhanced with complete code and settings upon possible acceptance.

Dependencies

our code is based on Python 3.8. The required packages can be found in requirements.txt

Usage

For training the model on the Cora dataset:

python3 main.py --n-layers 128  --n-epochs 200 --lr 0.1   --dataset cora   --model MGC   --weight-decay 0.0003  --dropout1 0.4  --n-hid 32  --dropout2 0.4 --alpha 0.05

For large-scale dataset, such as MAG240M dataset, where 120M paper nodes with 1.2B citations are involved for the node calssification task, we need pre-process node features first by runing:

python 120MAGpre.py --pos 
python 120MAGpre.py 

Then, we can train the model based on pre-processed node features by run:

python3 mainbatch120M.py   --n-epochs 300 --lr 0.00003  --gpu 0  --model MGC   --weight-decay 0.00000  --dropout1 0.1 --dropout2 0.1  --n-hid 2048 --alpha 0.05

Note that, due to the file size of the MAG240M node feature matrix, some scripts may require up to 256GB RAM.