For knwoledge graph completion, it is very common to use margin-based ranking loss

In the paper:margin-based ranking loss is defined as

$$ \min \sum_{(h,l,t)\in S} \sum_{(h',l,t')\in S'}[\gamma + d(h,l,t) - d(h',l,t')]_+$$

Here $d(\cdot)$ is the predictive model, $(h,l,t)$ means a positive training instance, and $(h',l,t')$ means a negative training instance corresponding to $(h,l,t)$.

However, in the Andrew's paper, it defines

$$ \min \sum_{(h,l,t)\in S} \sum_{(h',l,t')\in S'}[\gamma + d(h',l,t') - d(h,l,t)]_+$$

It seems that they switch the terms $d(h',l,t')$ and $d(h,l,t)$.

My question is that

does it matter to switch $d(h',l,t')$ and $d(h,l,t)$? it's real strange definition. Thanks


1 Answer 1


In this paper, $d$ denotes "dissimilarity" which should be minimized for positive samples.

In this paper, $d$ ($g$ in the paper) denotes "similarity" which should be maximized for positive samples (or equivalently $-g\left(T^{(i)}\right)$ should be minimized)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.