Signed-GCN

This is the source code for the manuscript "On the modelling and impact of negative edges in graph convolutional networks for node classification".

Requirements

To install requirements:

pip install -r requirements.txt

Reproduce results for synthetic data

Create a folder to save the results

mkdir job_id{1..25}

We have 25 files of results, which correspond to 25 combinations of within and between degree for each scenario. We generate between and within-community edges varying in ${1, 3, 6, 12, 24}$. For example, job_id1 corresponds to the case when within degree = 1 and between degree =1; $\dots$; job_id25 corresponds to the case when within degree = 24 and between degree =24.

If we want to obtain the result for the case when within degree = 1 and between degree =1, we implement:

python synthetic_data.py 1 1 0.1 0.3 0 0 1 1 1 1 1

The above eleven parameters correspond to:

• unique number of random seeds (here we have 30 random seeds in total corresponding to 30 splits)
• number of runs in each split
• training set percentage
• validation set percentage (the remaining is test set)
• negative link noise percentage
• positive link noise percentage
• within degree of Group 1
• within degree of Group 2
• within degree of Group 3
• between degree of the whole graph
• job id

Reproduce results for real networks

Create a folder to save the results

mkdir result

For Cora data, if we want to obtain the result for the case when there is no edge noise and no feature noise, we run this command:

python real_network.py 1 1 0 0 0

The above five parameters correspond to:

• unique number of random seeds corresponding to different weight initializations (here we have 30 different weight initializations in total)
• number of runs
• negative link noise percentage
• positive link noise percentage
• feature noise