Class precision: huge difference between True Pos. and True Negative

Olus

I have run Auto-model for a Dataset for a predictive model on Binary Label: Positive or Negative. 
All of the models provide a satisfying class rate on True Pos  (+/-98%) but the True Neg. class recalls are incredibly low: 1-7%! How should I interpret this? Can I still use the model? 
Many thanks for your help

lionelderkrikor

Hi @Olus,

How should I interpret this?

Difficult to answer without seeing your data. But maybe your dataset is highly imbalanced.
It means that, for example, the Positive class of your label represent 90% of your dataset and the Negative class of your label 
represents 10% of your dataset.

Can I still use the model?

It depends of what is your final goal : 
 - if your goal is to correctly predict the Positive class of your label , yes in deed , you can use this model...
 - if your goal is to correctly predict the Negative class of your label , no it's not the right model...

To have a more relevant answer, you can share your data and explain what is your goal....


Regards,

Lionel

lionelderkrikor

Hi @Olus,

How should I interpret this?

Difficult to answer without seeing your data. But maybe your dataset is highly imbalanced.
It means that, for example, the Positive class of your label represent 90% of your dataset and the Negative class of your label 
represents 10% of your dataset.

Can I still use the model?

It depends of what is your final goal : 
 - if your goal is to correctly predict the Positive class of your label , yes in deed , you can use this model...
 - if your goal is to correctly predict the Negative class of your label , no it's not the right model...

To have a more relevant answer, you can share your data and explain what is your goal....


Regards,

Lionel

Olus

Hi Lionel,
Many thanks for your answer. Strange thing is that I have a balanced dataset with 69% of positives and 31% negatives. As I want to predict positives I I'll use it. 
Cheers,
Olus