OR, as Steve suggested in a previous post, would it make more sense in
training an SVM to convert a single nominal column into a series of
binary columns?
"....
color = ("red, "blue", "green")
So, imagine if the features for your examples were color and height,
your "feature matrix" for N examples would be N x 4
0,1,0,15 # blue object, height 15
1,0,0,10 # red object, height 10
0,0,1,5 # green object, height 5
...."
From my LIMITED knowledge, it seems like an SVM would be more accurate
with unique binary columns for each value of a nominal factor, but I'm
not sure.
Can anyone provide an opinion on this?
Thanks!
-N
On 8/12/09 2:21 PM, Achim Zeileis wrote:
> On Wed, 12 Aug 2009, Noah Silverman wrote:
>
>> Hi,
>>
>> The answers to my previous question about nominal variables has lead
>> me to a more important question.
>>
>> What is the "best practice" way to feed nominal variable to an SVM.
>
> As some of the previous posters have already indicated: The data
> structure for storing categorical (including nominal) variables in R
> is a "factor".
>
> Your comment about "truly nominal" is wrong. A character variable is a
> character variable, not necessarily a categorical variable.
> Categorical means that the answer falls into one of a finite number of
> known categories, known as "levels" in R's "factor" class.
>
> If you start out from character information:
>
> x <- c("red", "red", "blue", "green", "blue")
>
> You can turn it into a factor via:
>
> x <- factor(x, levels = c("red", "green", "blue"))
>
> R now knows how to do certain things with such a variable, e.g.,
> produces useful summaries or knows how to deal with it in regression
> problems:
>
> model.matrix(~ x)
>
> which seems to be what you asked for. Moreover, you don't need call
> this yourself but most regression functions in R will do that for you
> (including svm() in "e1071" or ksvm() in "kernlab", among others).
>
> In short: Keep your categorical variables as "factor" columns in a
> "data.frame" and use the formula interface of svm()/ksvm() and you are
> fine.
> Z
>
>
>> For example:
>> color = ("red, "blue", "green")
>>
>> I could translate that into an index so I wind up with
>> color= (1,2,3)
>>
>> But my concern is that the SVM will now think that the values are
>> numeric in "range" and not discrete conditions.
>>
>> Another thought would be to create 3 binary variables from the single
>> color variable, so I have:
>>
>> red = (0,1)
>> blue = (0,1)
>> green = (0,1)
>>
>> A example fed to the SVM would have one positive and two negative
>> values to indicate the color value:
>> i.e. for a blue example:
>> red = 0, blue =1 , green = 0
>>
>> Or, do any of the SVM packages intelligently handle this internally
>> so that I don't have to mess with it. If so, do I need to be
>> concerned about different "translation" of the data if the test data
>> set isn't exactly the same as the training set.
>> For example:
>> training data = color ("red, "blue", "green")
>> test data = color ("red, "green")
>>
>> How would I be sure that the "red" and "green" examples get encoded
>> the same so that the SVM is accurate?
>>
>> Thanks in advance!!
>>
>> -N
>>
>> ______________________________________________
>> R-help@r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-help
>> PLEASE do read the posting guide
>> http://www.R-project.org/posting-guide.html
>> and provide commented, minimal, self-contained, reproducible code.
>>
>>
[[alternative HTML version deleted]]
______________________________________________
R-help@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.
