Official Implementation of the paper "Toward Degree Bias in Embedding-Based Knowledge Graph Completion" (WWW 2023).
A fundamental task for knowledge graphs (KGs) is knowledge graph completion (KGC). It aims to predict unseen edges by learning representations for all the entities and relations in a KG. A common concern when learning representations on traditional graphs is degree bias. It can affect graph algorithms by learning poor representations for lower-degree nodes, often leading to low performance on such nodes. However, there has been limited research on whether there exists degree bias for embedding-based KGC and how such bias affects the performance of KGC. In this paper, we validate the existence of degree bias in embedding-based KGC and identify the key factor to degree bias. We then introduce a novel data augmentation method, KG-Mixup, to generate synthetic triples to mitigate such bias. Extensive experiments have demonstrated that our method can improve various embedding-based KGC methods and outperform other methods tackling the bias problem on multiple benchmark datasets.
All experiments were conducted using python 3.9.12.
For the required python packages, please see requirements.txt.
tqdm>=4.0
torch>=1.10
torch_geometric>=2.0
numpy>=1.21
tensorboard>=2.5
optuna>=2.10.0
matplotlib>=3.6.2
pandas>=1.4.4
First clone our repository and install the required python packages.
git clone https://github.com/HarryShomer/KG-Mixup.git
cd KG-Mixup
pip install -r requirements.txt
The code relies on the kgpy library. A version is found here in the kgpy directory.
It is necessary to install it as a python package. To do so cd into the directory and run:
cd kgpy
pip install -e .
For KG-Mixup, the embeddings are extracted from a pre-trained model. The scripts for pre-training each model are in the scripts/pretrain folder. For example, to pretrain ConvE on FB15K-237:
cd scripts/pretrain
bash conve_fb15k237.sh
Once training is done, the model will be saved in the checkpoints/DATASET folder. For ConvE on FB15K-237 it will saved as checkpoints/FB15K-237/conve_fb15k237_pretrain.tar.
The scripts for replicating KG-Mixup can be found in the scripts/kg_mixup folder. As a reminder, you must have already pre-trained the model. Below is how to replicate the results for ConvE on FB15K-237:
cd scripts/kg_mixup
bash conve_fb15k237.sh
@inproceedings{shomer2023toward,
title={Toward Degree Bias in Embedding-Based Knowledge Graph Completion},
author={Shomer, Harry and Jin, Wei and Wang, Wentao and Tang, Jiliang},
booktitle={Proceedings of the ACM Web Conference 2023},
pages={705--715},
year={2023}
}