Options

Nominal variable= dummy-coded variable?

frankiefrankie Member Posts: 26 Contributor II
edited November 2018 in Help
Was wondering about this the other day. If I select a variable to be nominal, and it has X categories, is this mathematically equal to having manually edited the variable into n-1 (or n) dummy categories?

That is, if I use a method that accepts inputs of nominal measures, is there any difference if the nominal variable has been split into a group of binomial variables or not?


-Frankie

Answers

  • Options
    wesselwessel Member Posts: 537 Maven
    Depends on the learn algorithm used.

    Theoretically you just need log_2(k).

    The k-1 is used in ANOVA, ANCOVA, Generalised Linear Models, etc.
    Either coded 1/0 (dummy coding), or 1/-1 (effect coding).

    Please read:
    www.cs.washington.edu/homes/jbigham/cv/princeton-thesis.pdf
    On Using Error-Correcting Codes and Boosting to Learn Multi-Class Classification Problems

    Error-Correcting Codes is a very important keyword.
  • Options
    frankiefrankie Member Posts: 26 Contributor II
    Thanks, I will read up on the subject. Interesting stuff.

    To clarify a bit more though. What I meant by my question... will RM do these conversion to dummy variables in the background when the nominal-type has been chosen?
  • Options
    wesselwessel Member Posts: 537 Maven
    frankie wrote:

    Thanks, I will read up on the subject. Interesting stuff.

    To clarify a bit more though. What I meant by my question... will RM do these conversion to dummy variables in the background when the nominal-type has been chosen?
    Using what learner?
  • Options
    frankiefrankie Member Posts: 26 Contributor II
    Any learner actually, my question is more concerned with how RM deals with a variable that has been marked as nominal.


  • Options
    wesselwessel Member Posts: 537 Maven
    frankie wrote:

    Any learner actually, my question is more concerned with how RM deals with a variable that has been marked as nominal.
    RM tried to stay as closely as possible to the algorithm presented in the paper that introduced the learning algorithm.
    So if you use for example the algorithm ID3, it does not make dummies at all, it calculates information gain directly on the original attributes.
  • Options
    landland RapidMiner Certified Analyst, RapidMiner Certified Expert, Member Posts: 2,531 Unicorn
    Hi,
    RapidMiner does not transform any nominal typed variables to numerical ones if the learner does need this! We are sticking to the believe that the user should be in full control over what the process does.
    In fact it would be easy adding this, but since the dummy encoding of nominal variables can have a major impact on the (prediction) performance, we won't do this implicitly. Instead we are presenting the user QuickFixes with which the user is capable of solving the problem with one click.

    Greetings,
      Sebastian
Sign In or Register to comment.