随机森林模型中使用递归特征消除的特征选择错误

Question

I have some hundred samples and I have already classified them into four different classes (clusters). Now, I'm interested in identifying the best set of genes that classify the samples into different classes.

I want to apply randomforest with recursive feature elimination and detect the genes (features). My data looks like below. Just posting some example data here.

Above data is just an example: My original data is in dataframe df with 100 samples in first column and 4 classes in second column and columns 3 to column 1002 there are in total 1000 genes with expression values.

I'm using the below code but I see there is an error.

library(caret)
library(mlbench)
library(Hmisc)
library(randomForest)

# define the control using a random forest selection function
control <- rfeControl(functions=rfFuncs, method="cv", number=10)

# run the RFE algorithm
results <- rfe(df[,3:1002], df[,2], sizes = df[,1:1002], rfeControl=control)

There is an Error: And I feel that I'm doing wrong somewhere.

Error in summary.connection(connection) : invalid connection

Here I'm giving the dput of the above data.

df <-structure(list(Samples = structure(c(1L, 8L, 9L, 10L, 11L, 12L, 
13L, 14L, 15L, 2L, 3L, 4L, 5L, 6L, 7L), .Label = c("Sample1", 
"Sample10", "Sample11", "Sample12", "Sample13", "Sample14", "Sample15", 
"Sample2", "Sample3", "Sample4", "Sample5", "Sample6", "Sample7", 
"Sample8", "Sample9"), class = "factor"), Class = structure(c(1L, 
2L, 3L, 1L, 2L, 4L, 2L, 1L, 1L, 4L, 1L, 3L, 4L, 1L, 1L), .Label = c("Class1", 
"Class2", "Class3", "Class4"), class = "factor"), Gene1 = c(1.030078784, 
0.944152632, 0.140700452, 0.013432323, 0.265233165, -0.084496727, 
4.835469554, 0.089434913, -0.433436179, 1.462895475, -0.116005356, 
1.007868422, 0.244881864, -1.495666899, 0.364368654), Gene2 = c(1.407236415, 
1.229003431, -0.322221459, -1.361955252, 0.310963955, 0.80115063, 
4.27765356, 0.872413223, -0.568249851, 1.187873069, -0.255284575, 
1.878058722, -0.767371822, -0.859697473, 0.057304769), Gene3 = c(0.200772234, 
-0.048349737, 1.224274924, 0.492396142, 0.500786902, -0.731802706, 
1.853246564, 1.611995455, 0.287088678, 0.509235514, 2.031735375, 
3.074950771, 2.069407179, 0.886158642, 1.736798303), Gene4 = c(1.23309207, 
1.321282889, 2.403301108, 0.748860637, 1.019200751, 1.393254607, 
2.667976275, 1.158136576, 1.89503732, 2.178257717, 0.747697632, 
2.834410716, 0.028594536, -0.411039831, 1.100167946), Gene5 = c(0.883005616, 
0.570786704, 0.72649548, 4.705893892, 0.086345885, 0.502530136, 
2.681497202, 0.640362079, 0.327319762, 2.086767741, 1.853085301, 
1.001799748, 0.126208601, 0.911621722, 0.671191951), Gene6 = c(2.590519025, 
3.076688902, 1.77414005, 1.014363629, 1.134652225, 2.71957962, 
4.696379063, -0.301828123, 1.214261665, 2.413881644, -0.470794827, 
0.520494891, 0.194511306, 0.075331863, 2.315680177), Gene7 = c(0.088929673, 
0.472549468, -0.125630236, -0.069648505, -0.715250242, 0.068554966, 
4.131662998, -0.075265565, -1.234425917, 0.343350342, 0.190414782, 
1.153495806, 0.210317581, -0.475603641, 0.294299351), Gene8 = c(2.112231178, 
2.780100532, 2.423828553, 1.569215682, 1.018119196, 2.583413401, 
6.483053565, 2.215201821, 1.893325529, 2.342058862, 4.001423142, 
4.221704757, 1.978211867, 1.452633851, 2.556589741)), class = "data.frame", row.names = c(NA, 
-15L))

Can anyone please tell me how I can use the above data and apply random forest to know which genes classify the samples into different classes. thanq.

Answer 1

sizes refer to the number of features you would like to try and retain, it should numeric but you provided something weird in df[,1:1002] .

See something like below, where i simulate a dataset and setting the sizes correctly ensures it runs along to choose the optimal number of features (from what you provide):

library(caret)
library(mlbench)
library(Hmisc)
library(randomForest) 

set.seed(101)
df = data.frame(samples=paste0("Samples",1:99),
                Class=paste0("Class",rep(1:3,33)),
                matrix(rnorm(99*1000),ncol=1000))

colnames(df)[3:ncol(df)]=paste0("Gene",1:1000)

# we create like 100 informative genes for Class1 and Class2
df[df$Class=="Class1",3:103] = df[df$Class=="Class1",3:103] + rpois(33*100,1.5)
df[df$Class=="Class2",104:203] = df[df$Class=="Class2",104:203] + rpois(33*100,1.5)

control <- rfeControl(functions=rfFuncs, method="cv", number=2)

# run the RFE algorithm
results <- rfe(df[,3:1002], df[,2], sizes = c(50,100,200), 
               rfeControl=control)

From the above, I ask for 50,100 or 200 informative features and I get:

results
Recursive feature selection

Outer resampling method: Cross-Validated (2 fold) 

Resampling performance over subset size:

 Variables Accuracy  Kappa AccuracySD KappaSD Selected
        50   0.9792 0.9688    0.02946 0.04419         
       100   0.9896 0.9844    0.01473 0.02210         
       200   1.0000 1.0000    0.00000 0.00000        *
      1000   1.0000 1.0000    0.00000 0.00000         

The top 5 variables (out of 200):
   Gene94, Gene198, Gene137, Gene136, Gene158

> results$optsize
[1] 200