Explanation post "Variable Importance in Prediction (Classification or Regression) Molels"

KNIME Analytics Platform
#1

I am trying to find a simple way to evaluate variable importance in a predictive model. I found this topic on the forum that give an example: https://tech.knime.org/forum/knime-general/variable-importance-in-prediction-classification-or-regression-molels 

Unfortunately, I don't really understand what the workflow is doing and why. Is there anyone who can give me a short explanation. What is the purpose of the target shuffling? And how can I evaluate which variables do matter and which variables are less important?

Anyone who can help?

Thanks in advance!

Wim

#2

Hi Wim,

I did an explanation here https://tech.knime.org/forum/knime-general/understanding-variable-importance

But I don't use any target shuffling in the workflow...

Best, Iris

 

#3

But, for a regression based model this is only true when the variables are independent, orthogonal, not when they are correlated. Otherwise, the joint effect(s) is not taken into account.

#4

Hi Farmer,

the idea behind is that if the model is worse without the column, this column is more important for the model. 

Shouldn't this work for joint effects as well? 

Best, Iris 

#5

Iris,

I'm afraid that unless the variables are independent (here meaning all correlations=0), it's not possible to isolate the contribution of each variable to the total variability( variance).

Best,

Farmer