This is the official implementation of "ReWise: A Relation-Wise Sampling Framework for Relational Graph Convolutional Networks." This paper was accepted for SEMANTICS 2024.
First, create an environment using the environment file provided, then activate it:
conda env create -f rewise-env.yml
conda activate rewise
Then install kgbench by downloading or cloning the kgbench repository [https://github.com/pbloem/kgbench-loader] and follow the installation steps or in the root
directory, run pip install .
Run the following to train and test the RGCN with ReWise-LDRN for amplus:
python main.py
For other datasets, hyperparameters, samplers, and settings specify the corresponding inputs. For example, to run ReWise-LDRN for dmgfull with multimodal features and sample size 64, run the following:
python main.py --data_name='dmgfull' --modality='all' --samp0=64
The input sampler, accepts the following options: LDRN, LDRE, LDUN, IDRN, IDUN, IARN, IAUN, and full-mini-batch. Set the batch_size to -1 to train with the full-batch setting.