Set Cutoff for Classification using Logistic Regression

btibert

Is it possible to manually set the threshold for the cutoff predict the label using a logistic regression?  I read that the cutoff is .5, which I get, but my dataset is heavily imbalanced and I would like to set this by hand.  There appears to be an automated way to do this, but for the sake of teaching the concept of the cutoff, I would prefer to show this manually.

Thanks!

arjun_gopal

Hi,
"Create Threshold" and "Apply Threshold" should do the trick for you. 

varunm1

Hello @btibert

The operators "create threshold" and "apply threshold" does this. Please inform if this is what you are looking for.

Hope this helps

MartinLiebig

Hi,
i think what @btibert relates to is Platt Scaling. The operator Rescale Confidences (Logistic) is i think what he looks for. You can combine this with Thresholds afterwards.

Cheers,
Martin

varunm1

@mschmitz oops my bad, I totally missed the Logistic Regression.

btibert

I had seen those operators @gopala but had a hard time wrapping my head around the construction, thanks for the screenshot as that is now intuitive.  Thanks!

Telcontar120

You can also use Drop Uncertain Predictions operator if you want to treat ambivalent cases as excluded rather than forcing them into one category or another simply by lowering (or raising) the threshold.  This is often another helpful way of dealing with the issue because it allows you to recalculate the performance metrics without the excluded cases.

btibert

Thanks @Telcontar120 , I will keep that in mind as well, but because this will likely be the first time my students have really sunk their teeth into logistic regression, the cutoff discussion, and modifying it manually is perfect for them to understand the construction before using tools that optimize it for them.  

Telcontar120

@btibert sure, that makes sense, but just for clarity, Drop Uncertain Predictions doesn't automatically optimize anything.  It simply excludes predictions below a certain confidence level that is set manually.  It is conceptually the same as Create Threshold, only Create Threshold says "use all data but don't change my prediction until the confidence is above 70%" and Drop Uncertain Predictions says "only keep predictions that are above 70% confidence."
If you take a look at the tutorial process it should make the outcome a bit clearer.

btibert

will do @Telcontar120 , many thanks for the follow-up note