TransE yield a comparable result with TransEA when training with TransEA hyperparameters

[phucty]

Dear Authors,

Thank you for sharing your source code.

I re-runned your implementation and it seems that numerical information does not help for the task of link prediction. Please take a look on the following result on FB15K. It seems that when using TransEA parameters, the TransE yield a slightly better result than TransEA.

	Raw		Filter
	MRR	H10	MRR	H10
TransE	0.1580	0.5031	0.2588	0.7030
TransEA	0.1662	0.5344	0.2996	0.7719
TransE with TransEA parameters	0.1667	0.5347	0.3000	0.7736

Could you explain how to reproduce the experiments with the claim in your paper?

Thank you very much.

Best, Phuc

[phucty]

Hi Olivier,

To run their code, I did edit the dataset location in initEA.cpp and init.cpp file string inPath = "./data/"; --> Change to string inPath = "./data/FB15K";

and then rebuild with bash makeEA.sh

Phuc

[phucty]

Hi Olivier, Yes, I still use this email. I am working on numerical data. If you interested, please contact me. Regards, Phuc

[kadimaolivier]

Hi Phuc,

Thank you very much,i really appreciate, i am very interested on numerical data,i am going to email you right now; once again thank you very much.

kind regards

Olivier