Split data set and low quality operator

AndyJ

Hi people,

I'm new to Rapidminer so be a bit passionate to me  

Two questions:
1. My issue is that I would like to split out my full data set into a training set(use this also for validation) and test set. From my understanding it's best practice to spilt out the test set straight away (and the before I do for instance any exploratory data analysis and feature selection analysis on the data). 

So If I split my full data set with "Split data" operator and use the "Remove correlated attributes" (referred to as "corr.") operator on the training set and the corr. operator remove some attributes. At the end I store this final training set. Now my test set has more attributes than my tranining set - how do I remove the same attributes generated by the corr. operator to my test set? I don't want to use the corr. operator on my test set because it could potensially remove fewer or other attributes. Is it possible to generation the test set in this automatic / dynamic way? Are there any other ways you guys do this process? 

2. Do there exists any "low quality" operator (i.e. the same low quality operations carried out inside the Turbo prep tab -> Cleanse -> Remove low quality) in Rapidminer Design? 

Love to hear from you. 

Best regards
Andy

Telcontar120

I would actually recommend cross validation instead of splitting your data into a single training vs testing sample.  This can be done using the cross-validation operator which has an excellent tutorial.
You can do your attribute removal either inside the cross-validation (technically correct) or outside the operator as part of your preprocessing (probably more common although it can bias your subsequent performance estimate).
Note that you don't actually need to worry about attribute removal per se: if you build a model on a subset of attributes and then apply it to a dataset that has extra attributes, those extra attributes will simply be ignored if you apply the same model.  But you do need to worry about attribute transformations, which is why you either need to do your data ETL beforehand (prior to all modeling) or do it inside the cross-validation and then pass all the transformations from the training data to the test data.