ParTrans-X

ParTrans-X is an efficient parallel lib for translating embedding algorithms.

The main adavantage of our lib is that our algorithms are totally lock-free multithreading, which enables the algorithms to be highly efficient, and we give some theoretical guarantees in our paper.

Up to now, we implemented the parallel version of TransE, AdagradTransE, and TransH. Named ParTransE, ParAdagradTransE, ParTransH respectively.

Also, we implemented the parallel version of the link prediction program for each of them, named ParTest_TransE and ParTest_TransH. (TransE and AdagradTransE use the same test program).

In addition, we provide all the shell scripts for parameter tuning.

Evaluation Results

We list the link prediction result and runtime of various methods implemented by ourselves in dateset FB15k and WN18.

FB15k,nepoch=1000(100 for AdagradTransE), threads=20

Model	MeanRank(Raw)	MeanRank(Filter)	Hit@10(Raw)	Hit@10(Filter)	Time(s)	Speedup Ratio
TransE	184	73	44.5	60.7	4658	-
ParTransE	185	69	45.3	62.3	496	9
ParAdagradTransE	186	70	44.9	61.9	42	111
TransH	183	60	46.6	65.5	6066	-
ParTransH	183	60	46.8	65.7	474	13

WN18,nepoch=1000(100 for AdagradTransE), threads=20

Model	MeanRank(Raw)	MeanRank(Filter)	Hit@10(Raw)	Hit@10(Filter)	Time(s)	Speedup Ratio
TransE	214	203	58.2	65.9	473	-
ParTransE	217	206	55.7	63.1	54	9
ParAdagradTransE	219	208	67.7	76.2	17	28
TransH	227	216	66.5	75.9	637	-
ParTransH	215	203	66.8	76.6	134	4.8

Compile and Use

Since we use openmp for multithreading, your g++ should contain it.

compile eg:

g++ -fopenmp ParTransE.cpp -o ParTransE

g++ -fopenmp ParTest_TransE.cpp -o ParTest_TransE

Use the script for parameter tuning is quite easy:

nohup ./train_test_TransE > logs 2>&1 &

More specific details can be found in the shell scripts, you can also make customized scripts.

Notice:

You should set $coreNum in the shell script less than the max number of your cores in your machine.

Cite

If you use the code, please kindly cite the following paper:

Denghui Zhang, Manling Li, Yantao Jia, Yuanzhuo Wang. Efficient Parallel Translating Embedding For Knowledge Graphs.IEEE/WIC/ACM International Conference on Web Intelligence 2017 (WI 2017).[pdf]

@inproceedings{zhang2017efficient,
  title={Efficient parallel translating embedding for knowledge graphs},
  author={Zhang, Denghui and Li, Manling and Jia, Yantao and Wang, Yuanzhuo and Cheng, Xueqi},
  booktitle={Proceedings of the International Conference on Web Intelligence},
  pages={460--468},
  year={2017}
}