I found it convenient in some cases to flatten categorical variables into
separate new columns instead of wrapping them into PooledDataArray or
CategoricalVariables. Here's some functions for doing this:
function getdummy{R}(df::DataFrame, cname::Symbol, ::Type{R})
darr = df[cname]
vals = sort(levels(darr))[2:end]
namedict = Dict(vals, 1:length(vals))
arr = zeros(R, length(darr), length(namedict))
for i=1:length(darr)
if haskey(namedict, darr[i])
arr[i, namedict[darr[i]]] = 1
end
end
newdf = convert(DataFrame, arr)
names!(newdf, [symbol("$(cname)_$k") for k in vals])
return newdf
end
function convertdummy{R}(df::DataFrame, cnames::Array{Symbol}, ::Type{R})
# consider every variable from cnames as categorical
# and convert them into set of dummy variables,
# return new dataframe
newdf = DataFrame()
for cname in names(df)
if !in(cname, cnames)
newdf[cname] = df[cname]
else
dummydf = getdummy(df, cname, R)
for dummyname in names(dummydf)
newdf[dummyname] = dummydf[dummyname]
end
end
end
return newdf
end
convertdummy(df::DataFrame, cnames::Array{Symbol}) = convertdummy(df,
cnames, Int32)
On Wednesday, October 1, 2014 11:24:58 PM UTC+3, Andrei Zh wrote:
>
> Thanks for your suggestions!
> So far using GLM and pool!() with a list of categorical variables works
> fine for me.
>
> On Wednesday, October 1, 2014 5:27:33 PM UTC+3, John Myles White wrote:
>>
>> Currently, the way to do this is via the GLM package (or at least its
>> strategy for generating design matrices), which handles indicators for you.
>>
>> Lots of improvements are possible, but we need better categorical data
>> support at a lower level before we can work on the improvements:
>> https://github.com/johnmyleswhite/CategoricalData.jl
>>
>> — John
>>
>> On Oct 1, 2014, at 6:22 AM, Stefan Karpinski <[email protected]>
>> wrote:
>>
>> Probably a better question for julia-stats, but there's also likely some
>> people here who can answer.
>>
>> On Wed, Oct 1, 2014 at 4:23 AM, Andrei <[email protected]> wrote:
>>
>>> Probably simple question, but I can't find any reference.
>>>
>>> What is the most convenient way to convert categorical variable to a set
>>> of dummy vars?
>>> So far I've been using
>>>
>>> int(indicatormat(df[:categoricalvar]))
>>>
>>> but it's pretty annoying to do it for every variable, give every
>>> indicator a name, resolve collinearity, etc.
>>> Is there any standard way to do these operations?
>>>
>>
>>
>>
