This repository is the official implementation of MStar, the method proposed in paper "Expanding the Scope: Inductive Knowledge Graph Reasoning with Multi-Starting Progressive Propagation".
python=3.7
conda create -n MStar python=3.7
conda activate MStar
conda install pytorch==1.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
pip install -r requirements.txt
pip install torch-scatter==2.0.8 -f https://data.pyg.org/whl/torch-1.8.0%2Bcu111.html
pip install torch-sparse==0.6.12 -f https://data.pyg.org/whl/torch-1.8.0%2Bcu111.html
- [-D] Dataset.
- [-T] Task, i.e., train/test.
- [-HW] Employ highway layer if selecting HW.
- [-E] Experiment name.
python3 train.py -D fb237_v1 -T train -HW --gpu 0 -E reproduction
python3 train.py -D fb237_v1 -T test -HW --gpu 0 -E reproduction
- [-M] Selection method. "None" removes entities selection. It works when not selecting HW.
- [--train_bad] Do not filter noisy samples if selecting train_bad.
python3 train.py -D fb237_v1 -T train --gpu 0 -M None -E wo_Selection
python3 train.py -D fb237_v1 -T train --gpu 0 -E wo_HighwayLayer
python3 train.py -D fb237_v1 -T train --gpu 0 -HW --train_bad -E wo_LinkVerify
Generate distance information for dataset fb237_v1
python3 analysis/dist_process.py -D fb237_v1
The distance information of fb237_v1
is output to analysis/dist_logs/dist_fb237_v1.log
.
The dataset fb237_v1
with distance for per-distance performance testing is output to data/fb237_v1_ind/test4.txt
.
Check per-distance performance by test and the result is output to test_results.txt
.
MStar is designed upon knowledge graph reasoning model RED-GNN. We thank them for making the code open-sourced.
@inproceedings{MStar,
title = {Expanding the Scope: Inductive Knowledge Graph Reasoning with Multi-Starting Progressive Propagation},
author = {Shao, Zhoutian and
Cui, Yuanning and
Hu, Wei},
booktitle = {ISWC},
year = {2024}
}