Talk:Knowledge graph embedding

Corrections in Definition section

• I removed the notation $(head,relation,tail)$ because it's not different from the already presented $(h,r,t)$
• The section "Models" should be revised thoroughly because it contains many more typos than I could fix in the time allotted

The article is actually pretty good, content-wise, and kudos to Edoardo for the length and the work! --Mirko Salaris (talk) 14:17, 18 June 2021 (UTC)

General comments

______

--Luthienrecanto (talk) 14:38, 18 June 2021 (UTC) Very comprehensive and detailed article. Good work!

• I changed a few typos.
• Also, I would suggest changing the order of the initial sentence to keep the standard format. So: KGE also referred to as ... in representation learning OR KGE in representation learning, also referred to as ...
• I also have a doubt about the first time you name "Rossi et al". Shouldn't it have the full names the first time you cite them? (Rossi, Barbosa, Firmani, Matinata, and Merialdo)

_______ Elisa191996 (talk) 15:00, 18 June 2021 (UTC)

• Nice article, good work! I made some changes but I couldn't be able to publish them. Probably other users were doing the same!
• just a question, I don't get why the application are in the middle of the article.

_______

The authors of the page should add RESCAL 2011 to the time line, as the first tensor based KGE! — Preceding unsigned comment added by Slloris (talk • contribs) 04:05, 4 October 2021 (UTC)

Hits@K formula seems incorrect

Hits@k formula is copied verbatim from the cited paper https://doi.org/10.1109%2FACCESS.2020.3030076 but I believe the formula does not seem to match the natural language description.

Hits@K$={\frac {|\{q\in Q:q<k\}|}{|Q|}}\in [0,1]$

"Given $Q$ as the set of all ranked predictions of a model"

Intuitively this formula reads as "the number of all predictions in the top k predictions" divided by "number of all predictions".

The natural language description says "H@K, is a performance index that measures the probability to find the correct prediction in the first top K model predictions" and this seems to be consistent over multiple sources (https://pykeen.readthedocs.io/en/stable/api/pykeen.metrics.ranking.HitsAtK.html ; https://doi.org/10.1109%2FACCESS.2020.3030076 ; https://arxiv.org/pdf/2003.08001).

Here is an alternative formula for this metric based on description, assuming the possibility of multiple correct answers:

For N model predictions, if we define the set $P\subseteq \{1,\ldots ,N\}$ as the set of the ranks for all the correct model predictions, the formular would be

Hits@K$={\frac {|\{p\in P:p<k\}|}{\min(N,k)}}$

Intuitively, this is "the number of true predictions in the top k predictions" divided by "the number of all predictions in the top k predictions" which is always k unless there are less than k predictions.

If we try to represent this formula using the notation in the article, with $Q$ as the set of the ranks for all predictions, we have:

$N=|Q|$

$P=\{q\in Q:{\text{isCorrectPrediction}}(q)\}$

Where isCorrectPrediction is an imaginary predicate that determines if rank q corresponds to a correct prediction. On the original formulation, this is not distinguished and that is what I believe is incorrect

And we can simplify the formula to:

Hits@K$={\frac {|\{q\in Q:q<k\land {\text{isCorrectPrediction}}(q)\}|}{\min(|Q|,k)}}$

Or alternatively, signaling the relation between divisor and dividend sets in formula:

Hits@K$={\frac {|\{q\in Q:q<k\land {\text{isCorrectPrediction}}(q)\}|}{|\{q\in Q:q<k\}|}}$

Which is different from the formula in the article.

I don't have time right now to properly research the matter at this point and come up with correct sources for the formula, but I submit this topic so that other contributors can check if the discrepancy is correct and potentially research it. Rafael-patronilo (talk) 10:49, 25 June 2026 (UTC)