search

ZichaoHuang / TransE

A TensorFlow implementation of TransE model
198 stars 62 forks source link

The result is too good to make me feel confused #12

Closed MGwave closed 5 years ago

MGwave commented 5 years ago

I set embedding size =300, margin=4, epoch =2000, and I got the following result : 188.4303 | 52.74941 | 0.518122 | 0.815788

It is too good. This version of TransE can beat most of the TransE-based methods. Why?

image

And, your code is important to me, thanks!

ZichaoHuang commented 5 years ago

From my point of view, the original TransE results (from Bordes's paper in NIPS 2013) are way below its true performance. Plus, the results shown in this repo are also not optimized. You can do a grid search to see how far it can get if you like. Feel free to ask any questions when checking the code.

MGwave commented 5 years ago

Thanks for your patience!