Relational Graph Convolutional Networks (RGCN) Pytorch implementation

Pytorch-based implementation of RGCN for semi-supervised node classification on (directed) relational graphs. The code is adpated from Kipf's Keras-based implementation. See details in Modeling Relational Data with Graph Convolutional Networks (2017) [1].

The code for the link prediction task in [1] can be found in the following repository: https://github.com/MichSchli/RelationPrediction

Dependencies

• Compatible with PyTorch 1.4.0 and Python 3.7.3.
• Dependencies can be installed using requirements.txt.

Environment

• The implementation is supposed to train in the GPU enviornment.
• I test all of the datasets with RGCN on GeForce RTX 2080 Ti and CPU with 128GB RAM.

Dataset:

• RGCN use AIFB, MUTAG, and BGS as benchmark datasets for semi-supervised node classification.
• AIFB, MUTAG, and BGS are included in data directory. The datasets are adapted from RDF2Vec (2016).

Training model (node classification):

We include early-stopping mechanisms in pytorchtools.py to pick the optimal epoch.

• Install all the requirements from requirements.txt.
• AIFB:

python run.py --data aifb --epochs 50 --bases 0 --hidden 16 --lr 0.01 --l2 0

• MUTAG:

python run.py --data mutag --epochs 50 --bases 30 --hidden 16 --lr 0.01 --l2 5e-4

• BGS:

python run.py --data bgs --epochs 50 --bases 40 --hidden 16 --lr 0.01 --l2 5e-4 --no_cuda

• AM:

python run.py --data am --epochs 50 --bases 40 --hidden 10 --lr 0.01 --l2 5e-4 --no_cuda

Note: Results depend on random seed and will vary between re-runs.

• --bases for RGCN basis decomposition
• --data denotes training datasets
• --hidden is the dimension of hidden GCN Layers
• --lr denotes learning rate
• --l2 is the weight decay parameter of L2 regularization
• --drop is the dropout value for training GCN Layers
• Rest of the arguments can be listed using python run.py -h