Building models with SAS Enterprise Miner, SAS Factory Miner, SAS Visual Data Mining and Machine Learning or just with programming

SEMMA

Accepted Solution Solved
Reply
Contributor
Posts: 45
Accepted Solution

SEMMA

Hi there

 

I may have what sounds like a stupid question but in SEMMA methodology, why sampling is first?

In other words, if I first manipulate my large data (imputing missing values/binning interval data etc...) and then after perform a sampling on this data to create my model is that complete non-sense?

Thanks

Nicolas 

 

 


Accepted Solutions
Solution
‎02-14-2018 04:56 AM
SAS Employee
Posts: 45

Re: SEMMA

Well, if there is one sample to do the analysis and another sample held out to evaluate the results of the analysis, and missing values are imputed using all the data, then the evaluation data is not completely independent if there are a bunch of missing values.  So, better practice, and practice simpler to explain and possibly avoid criticisms of the results,  is to impute & bin on each sample separately.   

That said, it often doesn't matter.  It's an art.

View solution in original post


All Replies
SAS Employee
Posts: 45

Re: SEMMA

Your approach is fine. I came up with "SEMMA" as an easily remembered guide for those who have little analytical experience.  People with analytical experience will do what they know best.

-Padraic

Contributor
Posts: 45

Re: SEMMA

Posted in reply to PadraicGNeville

Thank for your answer Padraic. The reason I asked is because I never came across (in my non-exhaustive search) work where the sampling was not performed straight on the raw imported full data-set. Nicolas

Solution
‎02-14-2018 04:56 AM
SAS Employee
Posts: 45

Re: SEMMA

Well, if there is one sample to do the analysis and another sample held out to evaluate the results of the analysis, and missing values are imputed using all the data, then the evaluation data is not completely independent if there are a bunch of missing values.  So, better practice, and practice simpler to explain and possibly avoid criticisms of the results,  is to impute & bin on each sample separately.   

That said, it often doesn't matter.  It's an art.

☑ This topic is solved.

Need further help from the community? Please ask a new question.

Discussion stats
  • 3 replies
  • 239 views
  • 0 likes
  • 2 in conversation