The code for our paper "Knowledge Graph Reasoning with Relational Digraph" which has been accepted by WebConf 2022.
A quick instruction is given for readers to reproduce the whole process.
Requirements
- pytorch 1.9.1+cu102
- torch_scatter 2.0.9
For transductive reasoning
cd transductive
python -W ignore train.py --data_path=data/WN18RR
For inductive reasoning
cd inductive
python -W ignore train.py --data_path=data/WN18RR_v1
We follow the rule mining methods, i.e., Neural-LP and DRUM, to randomly split triplets in the original train.txt
file into two files facts.txt
and train.txt
with ratio 3:1. This step is to make sure that the query triplets will not be leaked in the fact triplets used in RED-GNN. Empirically, increasing the ratio of facts, e.g. from 3:1 to 4:1, will lead to better performance.
Metrics | Family | UMLS | WN18RR | FB15k-237 | NELL-995 |
---|---|---|---|---|---|
MRR | .992 | .964 | .533 | .374 | .543 |
Hit@1 (%) | 98.8 | 94.6 | 48.5 | 28.3 | 47.6 |
Hit@10 (%) | 99.7 | 99.0 | 62.4 | 55.8 | 65.1 |
We use the full set of negative samples in evaluating the inductive results. This is different from the setting of 50 negative samples in GraIL.
metrics | WN-V1 | WN-V2 | WN-V3 | WN-V4 | FB-V1 | FB-V2 | FB-V3 | FB-V4 | NL-V1 | NL-V2 | NL-V3 | NL-V4 |
---|---|---|---|---|---|---|---|---|---|---|---|---|
MRR | .701 | .690 | .427 | .651 | .369 | .469 | .445 | .442 | .637 | .419 | .436 | .363 |
Hit@1 (%) | 65.3 | 63.3 | 36.8 | 60.6 | 30.2 | 38.1 | 35.1 | 34.0 | 52.5 | 31.9 | 34.5 | 25.9 |
Hit@10 (%) | 79.9 | 78.0 | 52.4 | 72.1 | 48.3 | 62.9 | 60.3 | 62.1 | 86.6 | 60.1 | 59.4 | 55.6 |
Readers are welcomed to fork this repository to reproduce the experiments and follow our work. Please kindly cite our paper
@inproceedings{zhang2022redgnn,
title={Knowledge graph reasoning with relational digraph},
author={Zhang, Yongqi and Yao, Quanming},
booktitle={Proceedings of the ACM Web Conference 2022},
pages={912--924},
year={2022}
}