Feature Interaction-aware Graph Neural Networks

This repository is the official implementation of Feature Interaction-aware Graph Neural Networks.

The sketches of FI-GNN model structure.

Requirements

To install requirements:

pip install -r requirements.txt

Training

You can train on four different datasets (BlogCatalog, Flickr, ACM, DBLP) by changing the parameters of data_name. You can also train three different GNNs (gcn, SimpleGCN, graphsage) by changing model_type

Semi-supervised Node Classification

• Training FI-GCN in node classification

cd FI_GCN
python train.py --supervised \
--epochs 200 \
--hidden 32 \
--lr 1e-3 \
--truncate_size 200 \
--model_type gcn \
--data_name BlogCatalog \
--output_file \path\to\output\file \
--weight_decay 5e-4 \
--dropout 0.2 

• Training FI-GraphSAGE in node classification

cd FI_GraphSAGE_SimpleGCN
python graphsage_classification.py \
--num-hidden 32 \
--lr 0.001 \
--dropout 0.2 \
--data_name BlogCatalog \
--output_file \path\to\output\file

• Training FI-SimpleGCN in node classification

cd FI_GraphSAGE_SimpleGCN
python simple_gcn_classification.py \
--num-hidden 32 \
--lr 0.001 \
--data_name BlogCatalog \
--output_file \path\to\output\file

Unsupervised Link Prediction

• Training FI-GCN in link prediction

cd FI_GCN
python train.py
--epochs 200 \
--hidden 32 \
--lr 1e-3 \
--truncate_size 200 \
--model_type gcn \
--data_name BlogCatalog \
--output_file \path\to\output\file \
--weight_decay 5e-4 \
--dropout 0.2 

• Training FI-GraphSAGE in link prediction

cd FI_GraphSAGE_SimpleGCN 
python link_prediction.py
--epochs 200 \
--hidden 32 \
--lr 1e-3 \
--truncate_size 200 \
--model_type graphsage \
--data_name BlogCatalog \
--output_file \path\to\output\file \
--weight_decay 5e-4 \
--dropout 0.2 

• Training FI-SimpleGCN in link prediction

cd FI_GraphSAGE_SimpleGCN 
python link_prediction.py
--epochs 200 \
--hidden 32 \
--lr 1e-3 \
--truncate_size 200 \
--model_type simplegcn \
--data_name BlogCatalog \
--output_file \path\to\output\file \
--weight_decay 5e-4 \
--dropout 0.2 

• Hyper-parameter search

We use grid search in our node classification task. You can check the parameter_search.sh as an example for gcn.

The hyper-parameters setting for these models are as follows:

Hyper-Parameter	Search Space
learning rate: lr	gcn:{0.01, 0.001, 0.005}, graphsage:{0.01, 0.001}, SimpleGCN:{0.2, 0.1}
hidden dimension: hidden	gcn, graphsage, SimpleGCN:{32, 32}
dropout ratio: dropout	gcn, graphsage, SimpleGCN:{0.2 0.5}
weight decay (l2 loss): weight_decay	gcn, SimpleGCN, graphsage:{5e-3, 5e-4}

Evaluation

By setting the output_file in the previous step, you can find the test result in \path\to\output\file.