Causal Inference Based Debiasing Framework for Knowledge Graph Completion

Code for Causal Inference Based Debiasing Framework for Knowledge Graph Completion

In this paper, we conduct a comprehensive analysis of these biases to determine their extent of impact. To mitigate these biases, we propose a debiasing framework called Causal Inference-based Debiasing Framework for KGC (CIDF) by formulating a causal graph and utilizing it for causal analysis of KGC tasks.

In this project, we employ CIDF on SimKGC.

Requirements

python=3.7
torch=1.11.0+cu113
transformers=4.27.1
wandb

All experiments are run with 4 V100(32GB) GPUs.

Download Datasets

For datasets WN18RR and FB25k-127, SimKGC provides their resources for downloading.

For datasets Wikidata5M, it can be downloaded as following:

bash ./scripts/download_wikidata5m.sh

The directory structure is as follows:

data
|- FB15k237
|- WN18RR
|- wikidata5m

Preprocess Datasets

After downloading the datasets, use the following bash commands to preprocess them.

bash scripts/preprocess.sh WN18RR

bash scripts/preprocess.sh FB15k237

bash scripts/preprocess.sh wiki5m_trans

bash scripts/preprocess.sh wiki5m_ind

Train and Evaluate

You can use the following bash commands to train and evaluate CIDF

OUTPUT_DIR=<model_saved_path> bash <script_path> <logger_name_of_wandb>

OUTPUT_DIR=./checkpoint/wn18rr/ bash scripts/train_wn.sh wn

OUTPUT_DIR=./checkpoint/fb15k237/ bash scripts/train_fb.sh fb

OUTPUT_DIR=./checkpoint/wikiT/ bash scripts/train_wiki_trans.sh wiki5m_trans

OUTPUT_DIR=./checkpoint/wikiI/ bash scripts/train_wiki_ind.sh wiki5m_ind

Evaluate Only

train_args.bin will be created in model_saved_path when training is finished.

You can reload the training set and evaluate the model as follow:

python evaluate.py -apath <model_saved_path>/train_args.bin

Citation

If you use our code in your research, please cite our work:

@inproceedings{10.1007/978-3-031-47240-4_18,
author = {Ren, Lin and Liu, Yongbin and Ouyang, Chunping},
title = {Causal Inference-Based Debiasing Framework for&nbsp;Knowledge Graph Completion},
year = {2023},
isbn = {978-3-031-47239-8},
publisher = {Springer-Verlag},
address = {Berlin, Heidelberg},
url = {https://doi.org/10.1007/978-3-031-47240-4_18},
doi = {10.1007/978-3-031-47240-4_18},
booktitle = {The Semantic Web – ISWC 2023: 22nd International Semantic Web Conference, Athens, Greece, November 6–10, 2023, Proceedings, Part I},
pages = {328–347},
numpages = {20},
keywords = {Knowledge Graph Completion, Causal Inference, Link Prediction},
location = {Athens, Greece}
}