anonymous

Preprocess

Construct a close world dataset Family and split the dataset in different integrity randomly. Run the following command.

bash init.sh

RuleTensor-TSP

python GraphRule.py -dataset=DATASET -rule_len=LEN -hc_thr=HC -sc_thr=SC -percent=PER -gpu=GPU

DATASET: choose the dataset in DATA/

LEN: set the length of rule

HC: set the head coverage threshold of rule

SC: set the standard confidence threshold of rule

PER: set the integrity of the dataset

GPU: -1 for cpu, otherwise the gpu id

HAKE-TSP

python runs.py -train -test -data=DATASET -gpu=GPU -perfix=PERFIX --valid_steps=STEP

PERFIX: set the integrity of the dataset in the format of percent_, like 0.6_

STEP: do valid every STEP steps

GPHT

1. generate subgraphs

python run.py -dataset=DATASET -subgraph=SUBLEN -perfix=PERFIX

SUBLEN: set max hops of subgraph from center to edge

2. pre-train embeddings

python run.py -dataset=DATASET -subgraph=SUBLEN -perfix=PERFIX -batch=BATCH -pretrain -desc=DESC

3. train the model

python run.py -dataset=DATASET -subgraph=SUBLEN -perfix=PERFIX -lr=LR -restore=RESTORE

LR: a little scale number for learning rate, like 0.00003 or less

4. predict triples(in HAKE-TSP)

python runs.py -train -test -data=DATASET -gpu=GPU -perfix=PERFIX --valid_steps=STEP -testGNN ../GPHT/EXPS/DATASET/toKGE_XXX.pt

Acknowledgement

We refer to the code of HAKE and CompGCN. Thanks for their contributions.