Generating synthetic data for unsupervised learning

1
votes

I want to prepare data for unsupervised learning with random forest. The procedure is as follows:

  • Take data and add attribute 'class' with value 1 to all examples
  • Generate synthetic data from original data:
    • while you don't have the same number of examples as in original data build examples:
      • sample new attribute value from all values of that attribute in original data
      • do that for all attributes and combine them into new example
  • assign to attribute 'class' of synthetic data value 2
  • bind both data together

At the end it look like this:

        ...      Class
                |1
     Original   |1
     Data       |1
                |1
    --------------
                |2
     Synthetic  |2
     Data       |2
                |2

My R code looks like this:

library(gtools) #for smartbind()

sample1 <- function(X)   { sample(X, replace=T) } 
g1      <- function(dat) { apply(dat,2,sample1) }

data$class <- rep(1, times=nrow(data)) #add attribute 'class' with value 1

synthData<-data.frame(g1(data[,1:ncol(data)])) #generate synthetic data with sampling from data
synthData$class <- rep(2, times=nrow(synthData)) #attribute 'class' is 2
colnames(synthData) <- colnames(data)
newData <- smartbind(data, synthData) #bind the data together

It's probably obvious that I'm really new to R, but it works - there is just one problem: types of attributes in synthetic data are not the same as in original data. If in original they are nums, now they become factors. How could I preserve same type while generating synthetic data?

Thank you!

Data1 (nums become factors):

structure(list(V2 = c(1.51793, 1.51711, 1.51645, 1.51916, 1.51131 ), V3 = c(13.21, 12.89, 13.44, 14.15, 13.69), V4 = c(3.48, 3.62, 3.61, 0, 3.2), V5 = c(1.41, 1.57, 1.54, 2.09, 1.81), V6 = c(72.64, 72.96, 72.39, 72.74, 72.81), V7 = c(0.59, 0.61, 0.66, 0, 1.76 ), V8 = c(8.43, 8.11, 8.03, 10.88, 5.43), V9 = c(0, 0, 0, 0, 1.19), V10 = c(0, 0, 0, 0, 0), realClass = structure(c(1L, 2L, 2L, 5L, 6L), .Label = c("1", "2", "3", "5", "6", "7"), class = "factor")), .Names = c("V2", "V3", "V4", "V5", "V6", "V7", "V8", "V9", "V10", "realClass"), row.names = c(27L, 138L, 77L, 183L, 186L), class = "data.frame")

Data2 (factors become chrs):

structure(list(realClass = structure(c(2L, 2L, 2L, 1L, 2L), .Label = c("e", "p"), class = "factor"), V2 = structure(c(6L, 3L, 4L, 6L, 6L), .Label = c("b", "c", "f", "k", "s", "x"), class = "factor"), V3 = structure(c(4L, 4L, 3L, 1L, 1L), .Label = c("f", "g", "s", "y"), class = "factor"), V4 = structure(c(5L, 5L, 5L, 3L, 4L), .Label = c("b", "c", "e", "g", "n", "p", "r", "u", "w", "y"), class = "factor"), V5 = structure(c(1L, 1L, 1L, 2L, 1L), .Label = c("f", "t" ), class = "factor"), V6 = structure(c(3L, 9L, 3L, 6L, 3L ), .Label = c("a", "c", "f", "l", "m", "n", "p", "s", "y" ), class = "factor"), V7 = structure(c(2L, 2L, 2L, 2L, 2L ), .Label = c("a", "f"), class = "factor"), V8 = structure(c(1L, 1L, 1L, 1L, 1L), .Label = c("c", "w"), class = "factor"), V9 = structure(c(2L, 2L, 2L, 1L, 1L), .Label = c("b", "n" ), class = "factor"), V10 = structure(c(1L, 1L, 1L, 10L, 4L), .Label = c("b", "e", "g", "h", "k", "n", "o", "p", "r", "u", "w", "y"), class = "factor"), V11 = structure(c(2L, 2L, 2L, 2L, 1L), .Label = c("e", "t"), class = "factor"), V12 = structure(c(NA, NA, NA, 1L, 1L), .Label = c("b", "c", "e", "r"), class = "factor"), V13 = structure(c(3L, 2L, 3L, 3L, 2L), .Label = c("f", "k", "s", "y"), class = "factor"), V14 = structure(c(3L, 3L, 2L, 3L, 2L), .Label = c("f", "k", "s", "y"), class = "factor"), V15 = structure(c(7L, 8L, 7L, 4L, 7L), .Label = c("b", "c", "e", "g", "n", "o", "p", "w", "y"), class = "factor"), V16 = structure(c(7L, 7L, 8L, 4L, 1L), .Label = c("b", "c", "e", "g", "n", "o", "p", "w", "y" ), class = "factor"), V17 = structure(c(1L, 1L, 1L, 1L, 1L ), .Label = "p", class = "factor"), V18 = structure(c(3L, 3L, 3L, 3L, 3L), .Label = c("n", "o", "w", "y"), class = "factor"), V19 = structure(c(2L, 2L, 2L, 2L, 2L), .Label = c("n", "o", "t"), class = "factor"), V20 = structure(c(1L, 1L, 1L, 5L, 3L), .Label = c("e", "f", "l", "n", "p"), class = "factor"), V21 = structure(c(8L, 8L, 8L, 4L, 2L), .Label = c("b", "h", "k", "n", "o", "r", "u", "w", "y"), class = "factor"), V22 = structure(c(5L, 5L, 5L, 5L, 6L), .Label = c("a", "c", "n", "s", "v", "y"), class = "factor"), V23 = structure(c(3L, 3L, 5L, 1L, 2L), .Label = c("d", "g", "l", "m", "p", "u", "w"), class = "factor")), .Names = c("realClass", "V2", "V3", "V4", "V5", "V6", "V7", "V8", "V9", "V10", "V11", "V12", "V13", "V14", "V15", "V16", "V17", "V18", "V19", "V20", "V21", "V22", "V23"), row.names = c(4105L, 6207L, 6696L, 2736L, 3756L), class = "data.frame")

1
rdatasetcluster-analysisrandom-forestunsupervised-learning
Since you don't show your data it's not obvious to see why you have factors in place of numbers, but you can do numcol <- as.numeric(as.character(factcol)) - dickoa
Yes, that works. Is there maybe a more general solution, so that no matter what type the attributes are, they stay the same after the procedure? - Uros K
It's easier to find answer with reproducible example. In this case we don't know much about your data (str(data) or better dput(data)). - dickoa
Thank's, meanwhile I post an answer. Check this out if it can solve your problem - dickoa

1 Answers

2
votes

You can always use this trick to have numeric columns

numcol <- as.numeric(as.character(factcol))

But I suspect that you have factor variable in your data.frame. Since apply return a matrix, if you have one factor in your data, all the numeric variable will be coerced to factor too.

Here is an example, using toy dataset 

set.seed(123)
toydat <- data.frame(A = 1:10, B = rnorm(10), C = LETTERS[1:10])
str(toydat)

## 'data.frame':    10 obs. of  3 variables:
##  $ A: int  1 2 3 4 5 6 7 8 9 10
##  $ B: num  -0.5605 -0.2302 1.5587 0.0705 0.1293 ...
##  $ C: Factor w/ 10 levels "A","B","C","D",..: 1 2 3 4 5 6 7 8 9 10

set.seed(1)
str(data.frame(apply(toydat[,1:2], 2, sample, replace = TRUE)))

## 'data.frame':    10 obs. of  2 variables:
##  $ A: num  3 4 6 10 3 9 10 7 7 1
##  $ B: num  1.5587 -0.2302 0.4609 0.0705 -1.2651 ...

# with the factor column C     
set.seed(2)
str(data.frame(apply(toydat[,1:3], 2, sample, replace = TRUE)))

## 'data.frame':    10 obs. of  3 variables:
##  $ A: Factor w/ 6 levels "10"," 2"," 5",..: 2 5 4 2 1 1 2 6 3 4
##  $ B: Factor w/ 8 levels " 0.129288","-0.230177",..: 8 7 6 2 1 5 3 7 1 4
##  $ C: Factor w/ 6 levels "B","D","E","G",..: 4 2 5 1 2 3 1 2 6 1

This is where the plyr package became useful, since you can control the output (using **ply). But in this case, the colwise function is sufficient

require(plyr)
set.seed(2)
mysamplingfun <- colwise(function(x) sample(x, replace = TRUE))
str(mysamplingfun(toydat[,1:3]))

## 'data.frame':    10 obs. of  3 variables:
##  $ A: int  2 8 6 2 10 10 2 9 5 6
##  $ B: num  1.715 1.559 -1.265 -0.23 0.129 ...
##  $ C: Factor w/ 10 levels "A","B","C","D",..: 7 4 9 2 4 5 2 4 10 2