Hierarchical Message-Passing Graph Neural Networks

Code of the first practical model (HC-GNN).

Required packages

The code has been tested running under Python 3.7.3. with the following packages installed (along with their dependencies):

• numpy == 1.16.5
• pandas == 0.25.1
• scikit-learn == 0.21.2
• networkx == 2.3
• community (python-louvain) == 0.13
• pytorch == 1.1.0
• torch_geometric == 1.3.2

Data requirement

All eight datasets we used in the paper are all public datasets which can be downloaded from the internet.

Code execution

Link prediction:

python main.py --task LP --dataset grid --mode basemodel --model HCGNN --layer_num 3 --epoch_num 2001 --lr 0.0001 --relu True --dropout True --drop_ratio 0.5 --same_level_gnn GCN --down2up_gnn MEAN --up2down_gnn GAT --fshot False --SEED 123 --gpu True

Node classification and community detection:

python main.py --task NC --dataset cora --mode basemodel --model HCGNN --layer_num 2 --epoch_num 201 --lr 0.01 --relu True --dropout False --drop_ratio 0.5 --same_level_gnn GCN --down2up_gnn MEAN --up2down_gnn GCN --fshot True --SEED 1234 --gpu True

Model hyper-parameters:

--task: the target downstream task, "LP; NC; Inductive", type=str, default=LP 
--dataset: dataset name, type=str, default=grid 
--mode: the experiment type, type=str, default=basemodel 
--model: the model name, type=str, default=HCGNN 
--layer_num: the number of layers of primary GNN encoder for within level propagation, type=int, default=3 
--epoch_num: epoch number, type=int, default=2001 
--lr: learning rate, type=float, default=0.0001 
--relu: whether use relu as activation function in the model, type=bool, default=True 
--dropout: whether use dropout component in the model, type=bool, default=True 
--drop_ratio: dropout ratio if use dropout component, type=float, default=0.5 
--same_level_gnn: the GNN encoder for within level propagation, type=str, default=GCN 
--down2up_gnn: define the down2up propagation, type=str, default=MEAN 
--up2down_gnn: define the top2down propagation, type=str, default=GAT 
--fshot: if adopt few-shot learning settings, type=bool, default=False 
--SEED: random seed, type=int, default=123 
--gpu: if use GPU device, type=bool, default=True

Two demo file is given to show the execution of link prediction (LP) and node classification (NC) tasks.

Cite

Please cite our paper if it is helpful in your own work:

@article{ZLP23,
author = {Zhiqiang Zhong and Cheng{-}Te Li and Jun Pang},
title = {Hierarchical Message-Passing Graph Neural Networks},
journal = {Data Mining and Knowledge Discovery (DMKD)},
volume = {37},
number = {1},
pages = {381--408},
publisher = {Springer},
year = {2023},
}