search

awslabs / dgl-ke

High performance, easy-to-use, and scalable package for learning large-scale knowledge graph embeddings.
https://dglke.dgl.ai/doc/
Apache License 2.0
1.28k stars 196 forks source link

How to interpret score value #155

Closed akastrin closed 4 years ago

akastrin commented 4 years ago

Suppose that one of the prediction algorithm returns the following values:

src  rel  dst   score
 1    0    12   -4.11393
 1    0    18   -6.10925
 1    0    13   -3.66778

I'm not quite sure how to interpret the score value: lower the score (more negative) greater the likelihood that prediction is true or vice versa? (So in the example above the relation 1-0-18 is more probable that relation 1-0-13?)

Thanks, Andrej

classicsong commented 4 years ago

The larger the score (closer to 0) the better. So in Your example, 1-0-13 is more likely to be connected than 1-0-12 and 1-0-12 is more likely than 1-0-18.

akastrin commented 4 years ago

Thanks!

FilipKon commented 3 years ago

The larger the score (closer to 0) the better. So in Your example, 1-0-13 is more likely to be connected than 1-0-12 and 1-0-12 is more likely than 1-0-18.

Hey,

but this is true if the scoring function is logsigmoid is that correct ? I'm training with RESCAL and if I create the prediction with and without the scoring function then I see that without the logsigmoid the bigger the number the better it is. And my other question would be if the exist min and max values for the predictions ?

Thanks. Best, Filip