MStar

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".

Requirements

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

Reproduction

• [-D] Dataset.
• [-T] Task, i.e., train/test.
• [-HW] Employ highway layer if selecting HW.
• [-E] Experiment name.

Train MStar

python3 train.py -D fb237_v1 -T train -HW --gpu 0 -E reproduction

Test MStar

python3 train.py -D fb237_v1 -T test -HW --gpu 0 -E reproduction

Ablation

• [-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

Per-distance Performance

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.

Acknowledgement

MStar is designed upon knowledge graph reasoning model RED-GNN. We thank them for making the code open-sourced.

Citation

@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}
}