How can I export conclusion from this confusion matrix table ?
I have 4 chest pain types , what is the actual correlation between reality and predictions? Should I take into account the recall or the precision?
Find more posts tagged with
Sort by:
1 - 2 of
21
Altair EmployeeHi @GeorgeOik1999 ,
this really really depends on your use case.
Best,
Martin
Hi!
The classical correlation is defined for two sets of numbers, it can't be calculated for this matrix. The accuracy (percentage of correct results) is what you get, and you are probably interested in the recall.
The importance of precision and recall (specificity and sensitivity) depends on the use case. Sometimes you're interested in a high recall for one class, sometimes in a high precision or the overall accuracy.
Here you have a very imbalanced data set with lots of ASY cases which skews the predictions into that direction. This results in a good recall for this class but bad values for the others. You might want to downsample that class, or use example weights so the models can predict the other classes better.
Regards,
Balázs
The classical correlation is defined for two sets of numbers, it can't be calculated for this matrix. The accuracy (percentage of correct results) is what you get, and you are probably interested in the recall.
The importance of precision and recall (specificity and sensitivity) depends on the use case. Sometimes you're interested in a high recall for one class, sometimes in a high precision or the overall accuracy.
Here you have a very imbalanced data set with lots of ASY cases which skews the predictions into that direction. This results in a good recall for this class but bad values for the others. You might want to downsample that class, or use example weights so the models can predict the other classes better.
Regards,
Balázs
Sort by:
1 - 1 of
11
Hi!
The classical correlation is defined for two sets of numbers, it can't be calculated for this matrix. The accuracy (percentage of correct results) is what you get, and you are probably interested in the recall.
The importance of precision and recall (specificity and sensitivity) depends on the use case. Sometimes you're interested in a high recall for one class, sometimes in a high precision or the overall accuracy.
Here you have a very imbalanced data set with lots of ASY cases which skews the predictions into that direction. This results in a good recall for this class but bad values for the others. You might want to downsample that class, or use example weights so the models can predict the other classes better.
Regards,
Balázs
The classical correlation is defined for two sets of numbers, it can't be calculated for this matrix. The accuracy (percentage of correct results) is what you get, and you are probably interested in the recall.
The importance of precision and recall (specificity and sensitivity) depends on the use case. Sometimes you're interested in a high recall for one class, sometimes in a high precision or the overall accuracy.
Here you have a very imbalanced data set with lots of ASY cases which skews the predictions into that direction. This results in a good recall for this class but bad values for the others. You might want to downsample that class, or use example weights so the models can predict the other classes better.
Regards,
Balázs