LRGNN -- Search to Capture Long-range Dependency with Stacking GNNs for Graph Classification

This repository is the code for our WWW 2023 paper: Search to Capture Long-range Dependency with Stacking GNNs for Graph Classification

Overview

In this paper, we provide a novel method LRGNN to capture the long-range dependencies with stacking GNNs in the graph classification task. We justify that the over-smoothing problem has smaller influence on the graph classification task, and then employ the stacking-based GNNs to extract the long-range dependencies. Two design needs, i.e., sufficient model depth and adaptive skip-connections, are provided when designing the stacking-based GNNs. To meet these two design needs, we unify them into inter-layer connections, and then design these connections with the help of NAS. Extensive experiments demonstrate the rationality and effectiveness of the proposed LRGNN.

Requirements

 torch-geometric==1.7.2
 torch-scatter==2.0.8
 torch==1.8.0+cu111
 numpy==1.18.5
 hyperopt==0.2.7
 python==3.8.3

Instructions to run the experiment

Step 1. Run the search process, given different random seeds. (The NCI1 dataset is used as an example)

(B8C1 Full) python train_search.py --data NCI1   --gpu 0 --num_blocks 8 --cell_mode full --num_cells 1 --agg gcn --cos_temp --BN

(B12C3 Repeat) python train_search.py --data NCI1   --gpu 0 --num_blocks 12 --cell_mode repeat --num_cells 3 --agg gcn --cos_temp --BN

(B12C3 Diverse) python train_search.py --data NCI1   --gpu 0 --num_blocks 12 --cell_mode diverse --num_cells 3 --agg gcn --cos_temp --BN

The results are saved in the directory exp_res, e.g., exp_res/nci1.txt.

Step 2. Fine tune the searched architectures. You need specify the arch_filename with the resulting filename from Step 1.

(B8C1 Full) python fine_tune.py --data NCI1 --gpu 0 --num_blocks 8 --num_cells 1 --cell_mode full   --hyper_epoch 20  --arch_filename exp_res/nci1.txt   --cos_lr --BN

Evaluation

To reproduce the SOTA performance in Table 2, please use the following code:

python reproduce.py --data NCI1  --gpu 0

Cite

Please kindly cite our paper if you use this code:

@inproceedings{wei2023search,
title={Search to Capture Long-range Dependency with Stacking GNNs for Graph Classification},
author={Wei, Lanning and He, Zhiqiang and Zhao, Huan and Yao, Quanming},
journal={WebConf},
year={2023}
}