Interpreting 'Weight By' Operator Results

B00100719

Hi - Simple question.  When interpreting 'Weight By' operator results, e.g Weight by SVM - is it the absolute value of the weight (i.e. ignoring the minus in negative weights) that should be considered or does, say, and attribute with a weight of -0.47 really have less value than one with a value of 0.31.  In short, should I be ignoring the minus sign?

varunm1

Hi  @B00100719

Can you take a look at below thread where there is a similar discussion about negative weights. I see that they both (+ and -) has their own significance based on the .

https://community.rapidminer.com/discussion/44650/negative-weights 

@lionelderkrikor can pitch in for this question I guess

Thanks
Varun

B00100719

Thank you - That thread seems to suggest the interpretation depends on the process and dataset.  I don't really understand.  Mine is a binomial classification problem and I would have thought the interpretation would be the same regardless. I'm still a little confused as to should I regard the heavily negative weighted attributes as significant or not!

B00100719

And another thing I don't understand.  How come when I run 'Weight by Rule' with the "Normalize Weights" checkbox ticked, I get results set that is in a completely different order to when I run it without that checkbox ticked?

MartinLiebig

@B00100719 ,

if you tick normalize, then the biggest weight is set to 1. You basically devide all weights by the maximum of all your weights. This can be useful for some filters if your weights are unnormalized by definition (like Chi-Square values)

BR,

Martin

B00100719

@mschmitz - My question was two fold.  1.  Should you take the absolute value?  2. Why has normalizing resulted in a different order of importance

Telcontar120

Yes, my understanding is that you should consider the absolute value of the weights to understand strength of effect.
I will defer to @mschmitz regarding the effect of the normalization option.  I agree that I would not expect it to change the order of the weights, unless it is doing the normalizing on the underlying attributes first and then the weighting, in which case this is feasible.