Suppose my dataset is a 100 x 3
matrix filled with categorical variables. I would like to do binary classification on the response variable. Let's make up a dataset with following code:
set.seed(2013)
y <- as.factor(round(runif(n=100,min=0,max=1),0))
var1 <- rep(c("red","blue","yellow","green"),each=25)
var2 <- rep(c("shortest","short","tall","tallest"),25)
df <- data.frame(y,var1,var2)
The data looks like this:
> head(df)
y var1 var2
1 0 red shortest
2 1 red short
3 1 red tall
4 1 red tallest
5 0 red shortest
6 1 red short
I tried to do random forest and adaboost on this data with two different approaches. The first approach is to use the data as it is:
> library(randomForest)
> randomForest(y~var1+var2,data=df,ntrees=500)
Call:
randomForest(formula = y ~ var1 + var2, data = df, ntrees = 500)
Type of random forest: classification
Number of trees: 500
No. of variables tried at each split: 1
OOB estimate of error rate: 44%
Confusion matrix:
0 1 class.error
0 29 22 0.4313725
1 22 27 0.4489796
----------------------------------------------------
> library(ada)
> ada(y~var1+var2,data=df)
Call:
ada(y ~ var1 + var2, data = df)
Loss: exponential Method: discrete Iteration: 50
Final Confusion Matrix for Data:
Final Prediction
True value 0 1
0 34 17
1 16 33
Train Error: 0.33
Out-Of-Bag Error: 0.33 iteration= 11
Additional Estimates of number of iterations:
train.err1 train.kap1
10 16
The second approach is to transform the dataset into wide format and treat each category as a variable. The reason I am doing this is because my actual dataset has 500+ factors in var1 and var2, and as a result, tree partitioning will always divide the 500 categories into 2 splits. A lot of information is lost by doing that. To transform the data:
id <- 1:100
library(reshape2)
tmp1 <- dcast(melt(cbind(id,df),id.vars=c("id","y")),id+y~var1,fun.aggregate=length)
tmp2 <- dcast(melt(cbind(id,df),id.vars=c("id","y")),id+y~var2,fun.aggregate=length)
df2 <- merge(tmp1,tmp2,by=c("id","y"))
The new data looks like this:
> head(df2)
id y blue green red yellow short shortest tall tallest
1 1 0 0 0 2 0 0 2 0 0
2 10 1 0 0 2 0 2 0 0 0
3 100 0 0 2 0 0 0 0 0 2
4 11 0 0 0 2 0 0 0 2 0
5 12 0 0 0 2 0 0 0 0 2
6 13 1 0 0 2 0 0 2 0 0
I apply random forest and adaboost to this new dataset:
> library(randomForest)
> randomForest(y~blue+green+red+yellow+short+shortest+tall+tallest,data=df2,ntrees=500)
Call:
randomForest(formula = y ~ blue + green + red + yellow + short + shortest + tall + tallest, data = df2, ntrees = 500)
Type of random forest: classification
Number of trees: 500
No. of variables tried at each split: 2
OOB estimate of error rate: 39%
Confusion matrix:
0 1 class.error
0 32 19 0.3725490
1 20 29 0.4081633
----------------------------------------------------
> library(ada)
> ada(y~blue+green+red+yellow+short+shortest+tall+tallest,data=df2)
Call:
ada(y ~ blue + green + red + yellow + short + shortest + tall +
tallest, data = df2)
Loss: exponential Method: discrete Iteration: 50
Final Confusion Matrix for Data:
Final Prediction
True value 0 1
0 36 15
1 20 29
Train Error: 0.35
Out-Of-Bag Error: 0.33 iteration= 26
Additional Estimates of number of iterations:
train.err1 train.kap1
5 10
The results from two approaches are different. The difference is more obvious as we introduce more levels in each variable, i.e., var1
and var2
. My question is, since we are using exactly the same data, why is the result different? How should we interpret the results from both approaches? Which is more reliable?