[R] Categorical Variables and Machine Learning

[Andrew Ziem]

Try the function ctree() in the package party or earth() in earth.   You can use factor variable as is, or you can transform the factor to binary variables (i.e., is_P is 0 or 1, is_D is 0 or 1).  In the second case, you can use any algorithm, and earth() automatically transforms factors to binary features.

However, you may find 120 variables is not much data.


Andrew



-----Original Message-----
From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org] On Behalf Of Lorenzo Isella
Sent: Thursday, February 17, 2011 7:14 AM
To: r-help
Subject: [R] Categorical Variables and Machine Learning

Dear All,
Please consider a dataframe like the one below (I am showing only a few
rows).

>          role degree strength weight count disparity intermittency
>            P     10       82  18017     2  2.317073  5.550314e-05
>            P      7      529   4345    60  5.178466  6.904488e-03
>            P      8      609   4382    10  6.204535  1.141031e-03
>            D     42      230   6910    88  1.791153  6.367583e-03

You have a categorical variable (the role variable) which can assume
only a few values ("P","D","C","N","A") referring to different
individuals for whom you collect some extra properties (namely, degree,
strength, weight, disparity and intermittency, like in the table above).
My goal is to find the most suitable property (or combination of
properties) to guess the role of an individual. It looks like a typical
machine learning problem, but I have categorical variables to predict.
I am drowning in the wealth of R packages for machine learning, but I
really would like something simple and easy to use (consider that the
dataset covers only 120 individuals, so performance is not a problem).
Any suggestion is appreciated.
Cheers

Lorenzo